The Vulnpocalypse: AI Attack Harnesses and the Race to Reduce Exposure
AI can compress the distance between discovery and action, but more vulnerability findings do not automatically mean more successful attacks. The real question is whether organisations can find and reduce their exposed, exploitable risk faster than adversaries can act.
The vulnerability apocalypse is a pressure problem
Vulnpocalypse is shorthand for a feared surge in vulnerability discovery and exploitation as AI makes security research and repetitive investigation cheaper. It is not a forecast that every vulnerability becomes an incident. A model may help a researcher understand code, a scanner identify candidates, or an operator interpret results; an agentic harness can connect those tasks with tools, memory, and decisions. Each connection reduces some manual effort, but adds its own limitations and failure modes.
The pressure comes from a mismatch in speed and scale. An organisation may have thousands of internet-facing assets, multiple software versions, third-party services, and unclear ownership. An attacker needs only one reachable path that works. Defenders must identify exposure, establish applicability, coordinate change, and verify remediation across the estate. Faster attacker workflows can narrow that response window even when the underlying vulnerability classes are familiar.
Do not confuse an increasing number of reported findings with a measured increase in compromise. A discovered flaw might be unreachable, difficult to exploit, dependent on unusual configuration, or already mitigated. Public dashboards and scan counts are incomplete. Track attempts, validated exploitation, and business impact separately; models and scanners also produce false positives.
From discovery to impact
- Surface discovery: an operator assembles an inventory of reachable services and identifies likely technologies.
- Hypothesis and validation: tools examine candidate weaknesses; a plausible finding remains a hypothesis until supported by reliable evidence.
- Initial access: a successful exploit or stolen credential crosses an access boundary. Many attempts fail here.
- Follow-on activity: an intruder may seek additional privileges, move between systems, persist, or access data.
- Outcome: disruption, data exposure, financial loss, or a blocked attempt depends on the environment and response.
AI can assist at more than one stage without operating the entire chain autonomously. A human may select targets, review output, and perform critical actions. Conversely, a model may decide some technical steps inside a larger human-directed operation. Telemetry rarely establishes exactly how much work the model performed; avoid presenting a tool name as proof of autonomy or impact.
ARTEX and agentic attack workflows
ARTEX illustrates the concern around a controller that brings together an AI model and security tools. In the supplied inventory, it is classified as confirmed and associated with reconnaissance, exploitation, and post-exploitation. The dataset describes activity involving several operators and organisations; those details are dataset reports, not independently verified measurements in this article. Exposure of a controller or its presence in an investigation alone does not establish that it caused every nearby intrusion.
Agentic frameworks can coordinate scanners and other utilities, carry context between tasks, and present an operator with a plan or next action. Forks, rebranding, shared libraries, and changes in model providers complicate attribution. A public repository, a framework's banner, or an AI API connection is not a reliable stand-alone signature. Investigate the sequence of actions and its effect on your assets rather than relying on one product name.
The same general-purpose and security-testing tools can support legitimate, authorised assessments. Listing a project does not imply its maintainers endorse misuse. Organisations should approve their own testing, control agent credentials and tool permissions, and treat model output as untrusted until checked.
Harnesses in the supplied attack evidence
The supplied snapshot labels 9 entries confirmed: ARTEX, Cairn, CyberStrikeAI, PentestGPT, PentAGI, OpenClaw, Hermes Agent, oh-my-claudecode, Strix. Its confirmed tier means the dataset reports some attack-related use; it does not mean every entry caused a breach. For example, the snapshot describes Strix in a discovery role and oh-my-claudecode in supporting infrastructure operations.
The snapshot labels 5 entries observed: HexStrike AI, DeepAudit, AntiFox, CyberStrike, HalberdStrike. These records carry important caveats: installation, infrastructure association, reported experimentation, and unverified outputs are not interchangeable with proof that a particular tool compromised an organisation. The table preserves each record's outcome and description. These 14 entries are all of the evidence-tier rows in this file, not all tools used by every threat actor worldwide.
Another 140 entries are catalogued as having no attack on record in the supplied snapshot. That is a statement about this dataset, not a safety certification or a claim that no misuse has occurred elsewhere. Repository popularity is neither evidence of malicious use nor a risk score.
Explore the harness inventory
Thanks to Daniel Miessler for the work behind the harness landscape and for helping make this rapidly changing field easier to understand. This inventory is the supplied CSV snapshot, displayed as provided rather than a live feed. Its dates, actor names, descriptions, stars, and outcome labels are claims in that snapshot and may be incomplete, stale, or independently unverified. Blank fields are shown as an em dash; they do not mean zero.
The chart counts inventory rows, not incidents, victims, or unique actors. Use the filters to compare evidence tiers and search across displayed fields. uses seen in attacks describes reported use; uses from readme describes claimed capability and is not attack evidence. The source identifier column is omitted from the article to keep it free of references; the complete original CSV is available as a separate download.
Download the complete supplied CSV (154 records). The file includes source identifiers and repository addresses from the original dataset.
Showing all 154 records.
| name | table | tier | repo | kind | stars | license | last push | uses seen in attacks | uses from readme | actors | breach | description |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ARTEX | evidence | confirmed | — | — | — | — | — | recon exploit postex | — | VAPOR PANDA; Storm-0940; Storm-3358; CRAFTY FALCON; Unnamed actor; Onyx Sleet; Pearl Sleet | Yes | One controller hit 200+ organizations in 11 days, with access confirmed at nine. Storm-0940 used it to try sign-ins at 2,250+ Microsoft 365 tenants. Zscaler tied a theft of personal data on about 67,000 people from South Korean financial firms to it. Check Point read operator consoles: one operator stole 300+ iOS code-signing certificates. GMO found it on the attack infrastructure behind breaches of Japanese companies. Microsoft names two North Korean actors using it. |
| Cairn | evidence | confirmed | https://github.com/oritera/Cairn | — | 5301 | — | — | recon exploit postex | — | CL-STA-1238; CRAFTY FALCON; Retail card-theft operator; AIA0001; AIA0002 | Yes | Given a domain and a one-line goal, ran 25 operations against at least 19 organizations, with access or data taken at 10. Gambit counted 105 attack projects launched between 10 and 15 September and at least 27 companies compromised; NorForge found 20+ exposed panels and confirmed a live uploader web shell at a hardware-parts vendor. Google separately saw it used against a US aerospace target. |
| CyberStrikeAI | evidence | confirmed | https://github.com/AIPentest/CyberStrikeAI | — | 8770 | — | — | postex recon vuln exploit | scan | Unattributed operators; CRAFTY FALCON; Unattributed operator (Japan) | Yes | Implant deployed on education-sector networks in Latin America and Africa. The operator set the plan; the model on the command server chose and issued the technical steps: credential dumping, lateral movement, DCSync and Golden Ticket attacks, then database export. TrendAI separately confirmed it running 44 agents against at least 19 organizations, 16 or more in Japan, in October; deep compromise reached only a small share. Team Cymru found its banner on an IP Amazon had published from an AI-augmented actor's infrastructure, with NetFlow to the FortiGate devices that actor targeted. |
| PentestGPT | evidence | confirmed | https://github.com/GreyDGL/PentestGPT | — | 16468 | — | — | scan exploit | recon vuln research | UAT-10147; Unnamed operator | Yes | Installed on the attacker’s command server, scanned web servers and broke into at least one site. An operator running a rebuilt fork compromised at least 15 Argentine healthcare organizations. |
| PentAGI | evidence | confirmed | https://github.com/vxcontrol/pentagi | — | 34426 | — | — | scan exploit postex | — | GTG-50020; GTG-50029; CRAFTY FALCON | Yes | Named by Anthropic as a framework both groups used. One stole production AI API keys; the other dumped European political databases. Google saw most PentAGI activity stop at scanning, but some reach intrusion and persistence at multiple organizations. |
| OpenClaw | evidence | confirmed | https://github.com/openclaw/openclaw | — | 394672 | — | — | exploit postex | — | Unnamed operator; “Recon” operator | Yes | An agent on OpenClaw exploited a Telegram Mini App, dumped its database of 1,300+ users and verified the app's bot token. Google also found an .openclaw/ directory on a credential-harvesting server holding 23,800+ secrets. |
| Hermes Agent | evidence | confirmed | https://github.com/NousResearch/hermes-agent | — | 264173 | — | — | recon exploit scan postex | — | knaithe / KnYuan; Retail card-theft operator; Silent Lynx; BOSSAI operator | Yes | Four operations. Unit 42: exploit attempts against Langflow and n8n failed and a Citrix exploit run by hand got in. Gambit: the operator's console in a retail card-theft campaign. Acronis: the main attack agent in a lone operator's 12-agent mesh, extracting data at a Thai university. CloudSEK: handled recon, collection and log cleanup for an espionage cluster inside Tajikistan’s railway network. |
| oh-my-claudecode | evidence | confirmed | https://github.com/Yeachan-Heo/oh-my-claudecode | — | 40899 | — | — | scan | — | Zerofot | Tooling | Ran the IT and DevOps for a credential harvester its operator built with Codex and Claude Code: brought up, restarted and debugged the live scanner and its proxies. |
| Strix | evidence | confirmed | https://github.com/usestrix/strix | — | 76297 | — | — | vuln scan | recon exploit exploitdev postex | Suspected China-linked actor; Retail card-theft operator | Discovery only | In GTIG's suspected China-linked campaign, validated vulnerabilities. A honeypot caught it aimed at a live French shop with no permission prompts. Gambit saw it run 146 times against 138 hosts as the vulnerability-search stage of a retail card-theft operation, and some of its reports opened the next stage of exploitation, handed to Cairn. |
| HexStrike AI | evidence | observed | https://github.com/0x4m4/hexstrike-ai | — | 13694 | — | — | recon scan | exploitdev research | Suspected China-linked actor; BOSSAI operator | Not stated | Moved between recon tools on its own against a Japanese tech firm and an East Asian security company. Google listed it among frameworks abused in incident response in October, then described what it had seen of HexStrike as experimentation in November. Acronis found it installed as an MCP tool server on a lone operator's attack node and wired into every agent that needed it; the report ties no breach to it. |
| DeepAudit | evidence | observed | https://github.com/lintsinghua/DeepAudit | — | 7317 | — | — | — | recon scan vuln exploitdev | UAT-10147 | Not stated | Installed on the attacker's management server, and named among the frameworks the actor used to automate intrusions. Talos did not see it used against victims. |
| AntiFox | evidence | observed | https://github.com/AntiFox-AI/antifox | — | 748 | — | — | exploit postex | — | Unattributed AntiFox operators | Associated | Infrastructure hosting AntiFox on Alibaba Cloud was “associated with” confirmed intrusions: a web exploit, an IIS web shell, AWS credential checks and a successful read of a production secret. Arctic Wolf ties the activity to the infrastructure, not to the tool itself. |
| CyberStrike | evidence | observed | https://github.com/CyberStrikeus/CyberStrike | — | 3616 | — | — | vuln exploit | recon | BOSSAI operator | Not stated | One agent in a lone operator's 12-agent attack mesh. Acronis says it appears to have run vulnerability assessments and exploitation attempts against live targets, but its outputs could not be verified. |
| HalberdStrike | evidence | observed | https://github.com/kuaiyelink/HalberdStrike | — | 6 | — | — | recon vuln | — | BOSSAI operator | Not stated | Mapped the attack surface across at least 20 projects, from a Taiwanese emergency-services portal to 50+ commercial sites, and passed its reports to other agents for exploitation. Over half its high and critical findings were false positives or misclassifications. |
| Shannon | field | no attack on record | https://github.com/KeygraphHQ/shannon | Pentest agent | 52013 | AGPL-3.0 | 2026-11-16 | — | recon vuln exploit exploitdev | — | — | AI pentester for web apps and APIs. Reads the source, then attacks the running app to prove each finding. |
| reverse-skill | field | no attack on record | https://github.com/zhaoxuya520/reverse-skill | Skill pack | 44828 | MIT | 2026-11-14 | — | exploitdev research | — | — | Skill router that sends AI coding agents to reverse-engineering, pentest and CTF toolchains. |
| Agent Zero | field | no attack on record | https://github.com/agent0ai/agent-zero | General agent | 19900 | MIT | 2026-11-17 | — | — | — | — | General agent framework that writes and runs its own code and tools. |
| Claude-Red | field | no attack on record | https://github.com/SnailSploit/Claude-Red | Skill pack | 11173 | MIT | 2026-11-17 | — | recon vuln exploit exploitdev postex | — | — | Library of offensive security skills for the Claude skills system. |
| Claude-BugHunter | field | no attack on record | https://github.com/elementalsouls/Claude-BugHunter | Skill pack | 8917 | MIT | 2026-11-17 | — | recon vuln exploit postex | — | — | Claude Code skill bundle for bug hunting and external red-team work. |
| T3MP3ST | field | no attack on record | https://github.com/elder-plinius/T3MP3ST | Red-team platform | 7070 | AGPL-3.0 | 2026-09-08 | — | recon scan vuln exploit postex | — | — | Autonomous red-teaming platform: a multi-agent offensive meta-harness. |
| Decepticon | field | no attack on record | https://github.com/BitterSecurity/Decepticon | Red-team platform | 6634 | Apache-2.0 | 2026-11-17 | — | recon vuln exploit postex | — | — | Autonomous hacking agent for red teams. |
| Agentic-Bug-Hunter | field | no attack on record | https://github.com/awarexone/Agentic-Bug-Hunter | Bug-bounty agent | 5968 | MIT | 2026-11-12 | — | recon scan vuln exploit exploitdev | — | — | AI bug-bounty hunting toolkit. |
| METATRON | field | no attack on record | https://github.com/sooryathejas/METATRON | Pentest agent | 4281 | MIT | 2026-04-11 | — | recon scan vuln | — | — | Local-LLM pentest assistant that runs recon tools and has the model analyze the results. |
| RAPTOR | field | no attack on record | https://github.com/gadievron/raptor | Code auditor | 4032 | MIT | 2026-11-17 | — | vuln exploitdev | — | — | Claude Code set up as an offensive and defensive security research agent. |
| VulnClaw | field | no attack on record | https://github.com/Netw0rkNoob/VulnClaw | Pentest agent | 3937 | MIT | 2026-10-30 | — | recon vuln exploit exploitdev | — | — | Chinese-language pentest CLI agent that takes a target from recon through exploitation to a report. |
| ctf-skills | field | no attack on record | https://github.com/ljagiello/ctf-skills | Skill pack | 3730 | MIT | 2026-11-15 | — | research | — | — | Agent skills for solving CTF challenges across web, pwn, crypto, reversing and forensics. |
| RedAmon | field | no attack on record | https://github.com/samugit83/redamon | Pentest agent | 3428 | MIT | 2026-11-14 | — | recon scan exploit postex | — | — | Self-hosted pentest framework that maps the attack surface into a graph and works it autonomously. |
| PentestAgent | field | no attack on record | https://github.com/GH05TCREW/pentestagent | Pentest agent | 3307 | MIT | 2026-11-09 | — | recon scan | — | — | Agent framework for black-box testing: bug bounty, red team and pentest. |
| Pentest-Swarm-AI | field | no attack on record | https://github.com/Armur-Ai/Pentest-Swarm-AI | Pentest agent | 3248 | AGPL-3.0 | 2026-11-13 | — | recon scan vuln exploit exploitdev | — | — | Swarm of agents that runs recon, classification and exploitation. |
| Claude-OSINT | field | no attack on record | https://github.com/elementalsouls/Claude-OSINT | Skill pack | 3067 | MIT | 2026-10-20 | — | recon | — | — | Eight Claude skills that prime the model with offensive recon methodology: asset discovery and exposure checks. |
| Vulnhuntr | field | no attack on record | https://github.com/protectai/vulnhuntr | Code auditor | 2824 | AGPL-3.0 | 2025-02-06 | — | vuln exploitdev | — | — | Zero-shot vulnerability discovery with LLMs: traces code paths to find remotely exploitable bugs. |
| hack-skills | field | no attack on record | https://github.com/yaklang/hack-skills | Skill pack | 2667 | MIT | 2026-09-13 | — | postex | — | — | Knowledge base of 102 hacking skills that AI agents load, covering web, AD, pwn and mobile. |
| pentest-ai-agents | field | no attack on record | https://github.com/0xSteph/pentest-ai-agents | Skill pack | 2455 | MIT | 2026-10-06 | — | recon scan vuln exploit exploitdev postex | — | — | Offensive-security subagents for Claude Code. |
| Open Kritt | field | no attack on record | https://github.com/Kritt-ai/open-kritt | Code auditor | 2430 | AGPL-3.0 | 2026-11-16 | — | vuln | — | — | Self-hosted platform that orchestrates AI agents to find and validate vulnerabilities in code. |
| Deep Eye | field | no attack on record | https://github.com/zakirkun/deep-eye | Pentest agent | 2375 | MIT | 2026-10-30 | — | vuln | — | — | Pentest tool that has LLMs generate payloads and scan targets for 50+ vulnerability types. |
| sapper | field | no attack on record | https://github.com/N1nEmAn/sapper-v2 | Pentest agent | 2305 | MIT | 2026-11-11 | — | research | — | — | Multi-agent pentest orchestration harness that won a Tencent Cloud attack-defense challenge qualifier. |
| emp3r0r | field | no attack on record | https://github.com/jm33-m0/emp3r0r | C2 with MCP | 2112 | GPL-3.0 | 2026-11-17 | — | postex | — | — | Linux-first post-exploitation C2 whose MCP server lets an AI harness operate it in natural language. |
| pentest-ai | field | no attack on record | https://github.com/0xSteph/pentest-ai | Pentest agent | 1841 | MIT | 2026-11-16 | — | vuln exploit | — | — | AI pentester that re-runs each exploit to prove the finding. |
| Burp AI Agent | field | no attack on record | https://github.com/six2dez/burp-ai-agent | MCP server | 1594 | MIT | 2026-10-21 | — | scan vuln | — | — | Burp Suite extension that adds AI scanners and an MCP bridge for AI agents. |
| Pentest Copilot | field | no attack on record | https://github.com/bugbasesecurity/pentest-copilot | Assistant | 1563 | MIT | 2026-08-14 | — | recon scan exploit | — | — | Browser-based AI assistant for ethical hacking. |
| NeuroSploit | field | no attack on record | https://github.com/JoasASantos/NeuroSploit | Pentest agent | 1538 | MIT | 2026-11-16 | — | recon scan vuln exploit exploitdev postex | — | — | AI-powered penetration testing framework. |
| LuaN1aoAgent | field | no attack on record | https://github.com/SanMuzZzZz/LuaN1aoAgent | Pentest agent | 1486 | AGPL-3.0 | 2026-09-15 | — | vuln exploit | — | — | Fully autonomous pentest agent built on graph-based reasoning. |
| PentesterFlow | field | no attack on record | https://github.com/PentesterFlow/agent | Pentest agent | 1457 | Apache-2.0 | 2026-08-31 | — | recon vuln | — | — | Terminal pentest agent with skills, a Burp bridge and coverage tracking that keeps the analyst in control. |
| recon-skills | field | no attack on record | https://github.com/uphiago/recon-skills | Skill pack | 1385 | MIT | 2026-11-07 | — | recon | — | — | Skill pack for agents covering external recon, web and API testing, validation and attack paths. |
| Xalgorix | field | no attack on record | https://github.com/xalgorix/xalgorix | Pentest agent | 1366 | Apache-2.0 | 2026-11-17 | — | recon scan vuln exploit | — | — | Pentest agent that exploit-verifies every finding across a 22-phase run. |
| hackingBuddyGPT | field | no attack on record | https://github.com/ipa-lab/hackingBuddyGPT | Research framework | 1333 | MIT | 2026-10-24 | — | scan exploit postex research | — | — | Research framework for building LLM hacking agents in 50 lines or less. |
| Dark-Moon | field | no attack on record | https://github.com/ASCIT31/Dark-Moon | Pentest agent | 1259 | GPL-3.0 | 2026-10-27 | — | recon scan vuln exploit | — | — | Autonomous pentesting with 50 specialist agents across web, API, cloud and Active Directory. |
| AutoCVE | field | no attack on record | https://github.com/larlarua/AutoCVE | Code auditor | 1229 | AGPL-3.0 | 2026-10-01 | — | vuln | — | — | Chinese-language agent platform that audits source code for vulnerabilities and drafts CVE reports. |
| VulnHunter (Capital One) | field | no attack on record | https://github.com/capitalone/VulnHunter | Code auditor | 1180 | Apache-2.0 | 2026-10-07 | — | vuln exploitdev | — | — | Agentic source-code hunter that reasons like an adversary and drops findings it cannot prove exploitable. |
| Nebula | field | no attack on record | https://github.com/berylliumsec/nebula | Assistant | 1163 | BSD-2-Clause | 2026-11-17 | — | scan | — | — | Cybersecurity harness for AI-assisted engagements. |
| Vigolium | field | no attack on record | https://github.com/vigolium/vigolium | Pentest agent | 1151 | MIT | 2026-11-16 | — | scan exploitdev | — | — | Vulnerability scanner with an agentic mode whose swarm picks modules and writes custom attack extensions. |
| codex-redteam-mode | field | no attack on record | https://github.com/chAng-L19/codex-redteam-mode | General agent | 1137 | MIT | 2026-10-09 | — | — | — | — | Opt-in red-team operation runtime for Codex that injects a jailbreak prompt into the system layer. |
| PentestCode | field | no attack on record | https://github.com/s0ld13rr/pentestcode | Pentest agent | 1108 | MIT | 2026-11-05 | — | recon scan vuln exploit postex exploitdev | — | — | Multi-agent pentest system that keeps engagement state between runs. |
| AIRecon | field | no attack on record | https://github.com/pikpikcu/airecon | Pentest agent | 1094 | MIT | 2026-11-12 | — | recon scan vuln exploit exploitdev | — | — | Autonomous cybersecurity agent built on a self-hosted LLM through Ollama. |
| Clearwing | field | no attack on record | https://github.com/Lazarus-ai-OSS/clearwing | Pentest agent | 1090 | MIT | 2026-11-05 | — | recon scan exploit exploitdev | — | — | Autonomous scanner and vulnerability hunter that runs Kali tools and builds exploits with human approval. |
| Cybermes | field | no attack on record | https://github.com/Zyrexnn/Cybermes | Bug-bounty agent | 1038 | Apache-2.0 | 2026-11-16 | — | recon scan vuln exploit exploitdev | — | — | Offensive security and bug-bounty agent framework built on Hermes Agent. |
| Z3r0 | field | no attack on record | https://github.com/yv1ing/Z3r0 | Red-team platform | 1025 | MIT | 2026-09-13 | — | vuln | — | — | AI red-team workbench with specialist agents, sandboxed tooling and replayable timelines. |
| Cain | field | no attack on record | https://github.com/cdxiaodong/cain-agent | Pentest agent | 953 | Apache-2.0 | 2026-10-14 | — | recon vuln scan | — | — | AI pentest engineer for real authorized assessments, with 46 tools and a cloud penetration module. |
| ctf-agent | field | no attack on record | https://github.com/verialabs/ctf-agent | CTF agent | 939 | MIT | 2026-03-29 | — | research | — | — | Multi-model swarm that solves CTF challenges autonomously; won first place at BSidesSF 2026. |
| dsh-redteam-model | field | no attack on record | https://github.com/SeaOf0/dsh-redteam-model | Skill pack | 897 | MIT | 2026-10-27 | — | vuln exploit | — | — | Ten red-team working modes and runtime plugins for the DeepSeek harness; README in Chinese. |
| IoTHackBot | field | no attack on record | https://github.com/BrownFineSecurity/iothackbot | Skill pack | 890 | MIT | 2026-06-01 | — | vuln recon | — | — | IoT security testing toolkit with Claude Code skills for vulnerability discovery. |
| audit | field | no attack on record | https://github.com/evilsocket/audit | Code auditor | 889 | MIT | 2026-06-10 | — | vuln exploitdev | — | — | Eight-stage Claude agent pipeline that hunts vulnerabilities and compiles proof-of-concept code. |
| Ars0n Framework v2 | field | no attack on record | https://github.com/R-s0n/ars0n-framework-v2 | Bug-bounty agent | 883 | GPL-3.0 | 2026-11-16 | — | recon scan | — | — | Bug bounty recon framework whose MCP server exposes data and scans to AI assistants. |
| MCP-Kali-Server | field | no attack on record | https://github.com/Wh0am123/MCP-Kali-Server | MCP server | 879 | MIT | 2026-03-17 | — | recon exploit | — | — | MCP setup that gives an AI agent a Kali Linux machine and its tools. |
| numasec | field | no attack on record | https://github.com/FrancescoStabile/numasec | Pentest agent | 842 | AGPL-3.0 | 2026-05-08 | — | recon scan | — | — | AI agent for cyber security work. |
| FofaMap | field | no attack on record | https://github.com/asaotomo/FofaMap | Recon agent | 769 | Apache-2.0 | 2026-11-02 | — | recon scan | — | — | Chinese tool that ties FOFA asset discovery, AI reflection and Nuclei validation into one workflow. |
| MetasploitMCP | field | no attack on record | https://github.com/GH05TCREW/MetasploitMCP | MCP server | 765 | Apache-2.0 | 2026-02-05 | — | scan exploit exploitdev postex | — | — | MCP server that lets an agent drive Metasploit. |
| StrikeAgent | field | no attack on record | https://github.com/Yean-Sec/StrikeAgent_AtkBrain-Flash | Pentest agent | 745 | AGPL-3.0 | 2026-11-13 | — | vuln | — | — | Chinese AI pentest platform with red-team re-rating and re-verification of findings, tested in CTF settings. |
| Muteki | field | no attack on record | https://github.com/FishCodeTech/muteki | Pentest agent | 687 | AGPL-3.0 | 2026-11-14 | — | exploit | — | — | Offensive security harness for automated penetration testing. |
| src-hunter-skill | field | no attack on record | https://github.com/MyuriKanao/src-hunter-skill | Skill pack | 656 | MIT | 2026-05-24 | — | vuln | — | — | Chinese Claude Code skill for SRC and bug bounty hunting with attack playbooks and payloads. |
| MCP for Security | field | no attack on record | https://github.com/cyproxio/mcp-for-security | MCP server | 632 | MIT | 2026-03-30 | — | recon scan | — | — | MCP servers for SQLMap, FFUF, Nmap, Masscan and other security tools, for AI pentest workflows. |
| communitytools | field | no attack on record | https://github.com/transilienceai/communitytools | Skill pack | 624 | MIT | 2026-07-29 | — | recon scan vuln exploit exploitdev postex | — | — | Claude Code skills, agents and commands for pentesting and bug bounty. |
| CHYing-agent | field | no attack on record | https://github.com/yhy0/CHYing-agent | CTF agent | 596 | MIT | 2026-04-25 | — | research | — | — | Agent from Tencent Cloud AI pentest hackathons, ninth in the first year; README in Chinese. |
| Cyber-AutoAgent | field | no attack on record | https://github.com/westonbrown/Cyber-AutoAgent | Pentest agent | 552 | MIT | 2025-11-29 | — | recon scan exploit exploitdev | — | — | AI agent for autonomous cyber operations. |
| Mitsu | field | no attack on record | https://github.com/ARCANGEL0/Mitsu | Pentest agent | 541 | MIT | 2026-11-17 | — | recon vuln postex | — | — | Pentest agent client that runs an authorized engagement from recon through post-exploitation with 165+ MCP tools. |
| Flounder | field | no attack on record | https://github.com/adshao/flounder | Bug-bounty agent | 527 | AGPL-3.0 | 2026-11-15 | — | vuln exploitdev | — | — | Turns coding agents into a security auditor for bug bounty work, with local exploit construction. |
| Blitz Strike | field | no attack on record | https://github.com/shinthink/blitzstrike | MCP server | 518 | MIT | 2026-11-04 | — | recon vuln exploit | — | — | MCP security toolbelt that maps attack surface, traces source-to-sink paths and verifies findings live. |
| AIDA | field | no attack on record | https://github.com/Vasco0x4/AIDA | Pentest agent | 515 | AGPL-3.0 | 2026-10-21 | — | recon scan vuln exploit exploitdev | — | — | Turns any LLM into an autonomous pentester inside a scope you set. |
| Mantishack | field | no attack on record | https://github.com/deonmenezes/mantishack | Code auditor | 507 | Apache-2.0 | 2026-10-03 | — | vuln | — | — | AI vulnerability pipeline with recon, attacker-simulation validation and gated exploitation over a codebase. |
| Zen-Ai-Pentest | field | no attack on record | https://github.com/SHAdd0WTAka/Zen-Ai-Pentest | Pentest agent | 506 | MIT | 2026-11-03 | — | recon scan exploit postex | — | — | Multi-agent pentest framework with automated vulnerability scanning. |
| AdStrike | field | no attack on record | https://github.com/capture0x/AdStrike | Pentest agent | 489 | MIT | 2026-10-17 | — | exploit postex | — | — | Active Directory attack framework with an MCP server so the user's own LLM host drives 53 tools. |
| Rogue | field | no attack on record | https://github.com/faizann24/rogue | Pentest agent | 483 | GPL-3.0 | 2025-06-18 | — | vuln recon | — | — | LLM web vulnerability scanner agent that generates test payloads and verifies findings. |
| Pentest Harness | field | no attack on record | https://github.com/S1N6H/pentest-harness | Pentest agent | 474 | MIT | 2026-10-07 | — | — | — | — | Self-hosted AI agent harness for authorized pentests, bug bounty, security labs and CTFs. |
| BoxPwnr | field | no attack on record | https://github.com/0ca/BoxPwnr | CTF agent | 471 | AGPL-3.0 | 2026-07-24 | — | research | — | — | Framework that benchmarks LLM agents solving HackTheBox, PortSwigger, CTF and other security labs. |
| Bug-Bounty-Agents | field | no attack on record | https://github.com/matty69v/Bug-Bounty-Agents | Skill pack | 470 | MIT | 2026-04-30 | — | recon vuln exploitdev postex | — | — | Drop-in agent personas for bug bounty recon, web hunting, PoC validation and reporting. |
| OpenHack | field | no attack on record | https://github.com/openhackai/OpenHack | Code auditor | 445 | MIT | 2026-10-17 | — | vuln | — | — | Open-model agent scanner that hunts, validates and exploit-tests vulnerabilities in a codebase. |
| cyber-harness | field | no attack on record | https://github.com/chainreactors/cyber-harness | Pentest agent | 422 | AGPL-3.0 | 2026-11-17 | — | recon scan vuln | — | — | Single-binary agent for pentest, red team and bug bounty. |
| offensive-claude | field | no attack on record | https://github.com/hypnguyen1209/offensive-claude | Skill pack | 421 | MIT | 2026-10-22 | — | recon vuln exploitdev postex | — | — | Spec-driven offensive framework for Claude Code with 32 kill-chain skills. |
| RedCell | field | no attack on record | https://github.com/martian56/redcell | Pentest agent | 414 | MIT | 2026-11-12 | — | scan vuln exploit postex | — | — | Agents run a penetration test end to end inside a Kali container. |
| BugTraceAI | field | no attack on record | https://github.com/BugTraceAI/BugTraceAI | Pentest agent | 389 | Apache-2.0 | 2026-10-19 | — | recon scan vuln exploitdev | — | — | Self-hosted web pentesting suite with agentic scanning. |
| Vibe Pentest | field | no attack on record | https://github.com/ok-helloworld/vibe-pentest | Pentest agent | 387 | AGPL-3.0 | 2026-09-21 | — | vuln | — | — | Chinese multi-agent black-box pentest tool that mimics a human expert, including business-logic flaws. |
| Halberd | field | no attack on record | https://github.com/vectra-ai-research/Halberd | Red-team platform | 375 | GPL-3.0 | 2026-04-08 | — | postex | — | — | Multi-cloud attack emulation tool with a Claude-powered agent that helps discover and execute attack paths. |
| pentest-skills | field | no attack on record | https://github.com/crazyMarky/pentest-skills | Skill pack | 368 | Apache-2.0 | 2026-06-04 | — | recon | — | — | Chinese pentest skill set for Claude Code and Gemini CLI that turns natural language into tool runs. |
| h1-brain | field | no attack on record | https://github.com/PatrikFehrenbach/h1-brain | MCP server | 360 | MIT | 2026-10-13 | — | recon | — | — | MCP server linking an AI assistant to HackerOne bug bounty history, scopes and reports. |
| RedAI | field | no attack on record | https://github.com/kpolley/redai | Pentest agent | 356 | MIT | 2026-05-05 | — | vuln exploitdev | — | — | Terminal workbench that finds vulnerabilities with AI, then validates them live with PoC scripts. |
| TARS | field | no attack on record | https://github.com/osgil-defense/TARS | Pentest agent | 351 | MIT | 2025-01-17 | — | scan | — | — | Agents that automate pentesting. |
| Pojun | field | no attack on record | https://github.com/chu0119/Pojun | Red-team platform | 349 | AGPL-3.0 | 2026-11-17 | — | recon vuln exploit postex | — | — | Chinese multi-agent platform that takes a one-line goal through recon, exploitation and post-exploitation to a report. |
| DorkAgent | field | no attack on record | https://github.com/yee-yore/DorkAgent | Recon agent | 341 | MIT | 2026-08-24 | — | recon | — | — | LLM agent that automates Google dorking for bug hunting and pentesting. |
| LeoAI | field | no attack on record | https://github.com/cha0upup/LeoAI | C2 with MCP | 333 | GPL-3.0 | 2026-11-17 | — | postex | — | — | Chinese post-exploitation management platform with platform-level and per-node AI copilots. |
| deadend-cli | field | no attack on record | https://github.com/straylabs-ai/deadend-cli | Pentest agent | 331 | AGPL-3.0 | 2026-08-07 | — | recon vuln exploit exploitdev | — | — | Self-hosted agentic pentest tooling for black-box testing, on any LiteLLM model. |
| Lyrie | field | no attack on record | https://github.com/OTT-Cybersecurity-LLC/lyrie-ai | Pentest agent | 329 | MIT | 2026-11-17 | — | recon scan | — | — | Pentest CLI that fingerprints and scans targets with AI-assisted exploit validation; also red-teams LLM endpoints. |
| Apex | field | no attack on record | https://github.com/pensarai/apex | Pentest agent | 327 | Apache-2.0 | 2026-11-17 | — | vuln | — | — | Terminal pentest agent that runs blackbox and whitebox tests with agent swarms in a Kali container. |
| HackSynth | field | no attack on record | https://github.com/aielte-research/HackSynth | Research framework | 324 | AGPL-3.0 | 2025-06-24 | — | research | — | — | LLM agent and evaluation framework for autonomous penetration testing. |
| Plamen | field | no attack on record | https://github.com/PlamenTSV/plamen | Code auditor | 316 | MIT | 2026-09-26 | — | vuln exploitdev | — | — | Claude Code and Codex agent swarm that audits smart contracts and node clients with PoC exploits. |
| red-run | field | no attack on record | https://github.com/blacklanternsecurity/red-run | Skill pack | 309 | AGPL-3.0 | 2026-11-12 | — | recon postex | — | — | Claude Code skills, MCP servers and agent teams for pentest phases from recon to privilege escalation. |
| BurpMCP-Ultra | field | no attack on record | https://github.com/Cy-S3c/BurpMCP-Ultra | MCP server | 309 | MIT | 2026-09-28 | — | scan vuln | — | — | MCP server giving AI agents 150 Burp Suite Professional tools for scanning, fuzzing and traffic analysis. |
| LazyOwn | field | no attack on record | https://github.com/grisuno/LazyOwn | C2 with MCP | 302 | AGPL-3.0 | 2026-11-17 | — | recon scan exploit exploitdev postex | — | — | Multi-operator C2 that exposes 153 MCP tools to AI agents. |
| Scan-Spring-GO | field | no attack on record | https://github.com/sspsec/Scan-Spring-GO | MCP server | 302 | MIT | 2026-08-29 | — | exploit | — | — | Go scanner and exploiter for Spring components that exposes all its modules to AI clients over MCP. |
| WolfShell | field | no attack on record | https://github.com/0x7556/wolfshell | C2 with MCP | 286 | MIT | 2026-08-16 | — | postex | — | — | ASPX WebShell and C2 manager with an MCP server and AI-assisted penetration. |
| vulnerator-ai | field | no attack on record | https://github.com/irraz/vulnerator-ai | Pentest agent | 285 | MIT | 2026-11-17 | — | scan | — | — | OODA-loop pentest agent on Hermes, with Kali and Playwright tools over MCP. |
| SecGPT | field | no attack on record | https://github.com/ZacharyZcR/SecGPT | Pentest agent | 285 | Apache-2.0 | 2023-11-15 | — | scan vuln research | — | — | AutoGPT-style LLM tool for pentesting, vulnerability scanning and code audit, with CTF aims. |
| raptor-loop-hunt | field | no attack on record | https://github.com/dinosn/raptor-loop-hunt | Code auditor | 278 | MIT | 2026-09-01 | — | vuln | — | — | Claude Code skill that loops a multi-altitude vulnerability hunt over a codebase. |
| BlacksmithAI | field | no attack on record | https://github.com/yohannesgk/blacksmith | Pentest agent | 269 | AGPL-3.0 | 2026-10-23 | — | recon vuln exploit postex | — | — | Multi-agent pentest framework with separate recon, vulnerability, exploit and post-exploit agents. |
| AutoPentest AI | field | no attack on record | https://github.com/bhavsec/autopentest-ai | MCP server | 251 | Apache-2.0 | 2026-02-22 | — | recon vuln exploit | — | — | MCP server where Scout, Analyzer, Exploiter and Reporter agents run web pentests against OWASP test cases. |
| Haiou Workshop | field | no attack on record | https://github.com/haiou-666/Haiou-workshop | Assistant | 237 | MIT | 2026-10-25 | — | research | — | — | Chinese AI workbench for reverse engineering, authorized pentests and CTF using Codex and Claude Code. |
| VulnBot | field | no attack on record | https://github.com/KHenryAegis/VulnBot | Research framework | 236 | MIT | 2026-11-08 | — | research | — | — | Multi-agent framework for autonomous penetration testing. |
| Watchtower | field | no attack on record | https://github.com/fzn0x/watchtower | Pentest agent | 223 | MIT | 2026-11-15 | — | recon scan | — | — | LangGraph pentest CLI where LLM agents orchestrate 28 security tools. |
| Claude-AD | field | no attack on record | https://github.com/ADScanPro/Claude-AD | Skill pack | 218 | MIT | 2026-08-24 | — | recon postex | — | — | Active Directory pentest playbook for Claude Code as skills, agents and slash commands. |
| web3-bug-bounty-hunting-ai-skills | field | no attack on record | https://github.com/awarexone/web3-bug-bounty-hunting-ai-skills | Skill pack | 192 | MIT | 2026-09-21 | — | vuln | — | — | Eighteen Claude Code skill files for smart contract bug bounty hunting. |
| AIxVuln | field | no attack on record | https://github.com/m4xxxxx/AIxVuln | Code auditor | 191 | MIT | 2026-04-25 | — | vuln exploitdev | — | — | Chinese LLM multi-agent system that audits code, then verifies vulnerabilities with PoCs in Docker sandboxes. |
| kali-pentest | field | no attack on record | https://github.com/x-glacier/kali-pentest | Skill pack | 187 | Apache-2.0 | 2026-06-25 | — | postex | — | — | Kali Linux pentest skill for Claude Code, OpenClaw and Hermes Agent, with 200+ CLI tools. |
| NYU CTF agents | field | no attack on record | https://github.com/NYU-LLM-CTF/nyuctf_agents | CTF agent | 181 | MIT | 2026-07-17 | — | research | — | — | Baseline and D-CIPHER LLM agents that solve CTF challenges for the NYU CTF Bench. |
| Zypheron | field | no attack on record | https://github.com/KKingZero/Zypheron-CLI | Pentest agent | 174 | MIT | 2026-10-25 | — | recon | — | — | Local-first security CLI with recon, scanning and AI-assisted operator workflows. |
| AI-OPS | field | no attack on record | https://github.com/antoninoLorenzo/AI-OPS | Assistant | 172 | MIT | 2026-11-08 | — | scan | — | — | Penetration testing assistant built on open-source LLMs. |
| CkSKILLS | field | no attack on record | https://github.com/zhaji2333/CkSKILLS | Skill pack | 159 | MIT | 2026-11-17 | — | vuln recon postex | — | — | Chinese skill library that turns researcher methodology into reusable agent skills for SRC bug hunting. |
| forefy/.context | field | no attack on record | https://github.com/forefy/.context | Skill pack | 156 | MIT | 2026-11-17 | — | recon exploit exploitdev | — | — | AI agent skills for security audits, pentests and bug bounty work, from recon to PoC generation. |
| agent-smith | field | no attack on record | https://github.com/0x0pointer/agent-smith | Pentest agent | 155 | AGPL-3.0 | 2026-11-14 | — | recon exploit postex | — | — | Pentest framework where skills teach the LLM a methodology and it invents the attacks itself. |
| cochise | field | no attack on record | https://github.com/andreashappe/cochise | Pentest agent | 153 | MIT | 2026-10-30 | — | exploit postex research | — | — | Autonomous assumed-breach penetration testing of Active Directory networks. |
| FuzzingBrain | field | no attack on record | https://github.com/fuzzingbrain/afc-crs-all-you-need-is-a-fuzzing-brain | Code auditor | 153 | Apache-2.0 | 2026-11-17 | — | vuln exploitdev | — | — | LLM system built on OSS-Fuzz that finds vulnerabilities, builds PoC inputs and patches; a DARPA AIxCC entry. |
| Pentest MCP | field | no attack on record | https://github.com/DMontgomery40/pentest-mcp | MCP server | 151 | AGPL-3.0 | 2026-10-30 | — | scan | — | — | Scoped MCP server giving assessment agents nmap, Nuclei, Hashcat and other tools for authorized tests. |
| Heimdall | field | no attack on record | https://github.com/QiantangCredit/heimdall-agent | CTF agent | 151 | MIT | 2026-10-29 | — | research | — | — | Self-hosted AI solver and toolkit for CTF and security labs; top 2 at two Chinese challenges. |
| Incalmo | field | no attack on record | https://github.com/cylabcyberautonomy/Incalmo | Pentest agent | 150 | MIT | 2026-11-11 | — | research | — | — | Autonomous LLM-driven network pentest tool for multistage attacks, built for a research paper. |
| Executioner | field | no attack on record | https://github.com/dbx0/executioner | Pentest agent | 147 | MIT | 2026-11-17 | — | recon scan exploit | — | — | Local workspace where teams of AI agents run an assessment against your tools and scope. |
| bloodhound_mcp | field | no attack on record | https://github.com/mwnickerson/bloodhound_mcp | MCP server | 144 | GPL-3.0 | 2026-08-19 | — | recon | — | — | MCP server that lets an LLM query BloodHound attack paths across Active Directory and Entra ID. |
| Phalanx | field | no attack on record | https://github.com/webxos/phalanx | Pentest agent | 142 | MIT | 2026-08-16 | — | recon exploit postex scan | — | — | Multi-agent pentest framework with recon, exploit and post-exploit agents and 58 built-in tools. |
| PentestAgent (nbshenxm) | field | no attack on record | https://github.com/nbshenxm/pentest-agent | Pentest agent | 139 | MIT | 2025-12-20 | — | recon vuln research | — | — | Research LLM framework that automates intelligence gathering, vulnerability analysis and exploitation stages. |
| claude-pentest | field | no attack on record | https://github.com/Stickman230/claude-pentest | Pentest agent | 139 | MIT | 2026-06-08 | — | exploit exploitdev postex | — | — | Claude Code plugin with 15 agents and Kali MCP integration for full pentest engagements. |
| gori | field | no attack on record | https://github.com/hahwul/gori | MCP server | 139 | Apache-2.0 | 2026-11-17 | — | scan | — | — | Terminal intercepting proxy and pentest toolkit whose MCP server lets AI agents drive the same engagement. |
| PentestAssistant | field | no attack on record | https://github.com/InfSecLabs/PentestAssistant | Pentest agent | 138 | Apache-2.0 | 2025-04-06 | — | vuln scan | — | — | LLM web pentest detector using function calling and multiple agents to run nmap and sqlmap. |
| MAPTA | field | no attack on record | https://github.com/arthurgervais/mapta | Research framework | 136 | MIT | 2025-08-28 | — | — | — | — | Multi-agent system for autonomous web application security assessment. |
| Moxy | field | no attack on record | https://github.com/matank001/Moxy | Pentest agent | 134 | Apache-2.0 | 2026-01-22 | — | recon | — | — | Intercepting-proxy DAST tool where an AI agent reasons about attack surface, edits traffic and drives a browser. |
| pentestkit | field | no attack on record | https://github.com/lordx64/pentestkit | Pentest agent | 127 | MIT | 2026-07-31 | — | recon exploit research | — | — | Multi-agent pentest framework on the Claude Agent SDK where a skeptical verifier tries to exploit each candidate. |
| Hexestra | field | no attack on record | https://github.com/ABSllk/Hexestra | Assistant | 119 | Apache-2.0 | 2026-11-17 | — | scan | — | — | AI-native pentest workbench where Claude Code or Codex works with browser, shells and traffic capture. |
| aimy-skill | field | no attack on record | https://github.com/Prohao42/aimy-skill | Skill pack | 115 | MIT | 2026-11-09 | — | exploit | — | — | Chinese pentest skill for AI agents with a structured check interface and exploit-chain engine. |
| Zack AI Scanner | field | no attack on record | https://github.com/ZackSecurity/Zack-AI-Scanner | Pentest agent | 112 | GPL-3.0 | 2026-09-27 | — | vuln exploit | — | — | Burp Suite extension that runs an LLM pipeline against one request and sends the real payloads. |
| ZeroTrace | field | no attack on record | https://github.com/sergio11/zero_trace_hacking_tool | Recon agent | 111 | MIT | 2025-12-17 | — | recon | — | — | Personal AI-powered dorking tool that uncovers sensitive information for recon. |
| awsome-kali-MCPServers | field | no attack on record | https://github.com/ccq1/awsome_kali_MCPServers | MCP server | 110 | Apache-2.0 | 2025-11-16 | — | scan | — | — | Collection of MCP servers that expose Kali Linux tools to AI clients. |
| AutoPentester | field | no attack on record | https://github.com/YasodGinige/AutoPentester | Pentest agent | 109 | MIT | 2026-09-25 | — | research | — | — | Tool for automated pentesting tasks. |
| CyberSentry | field | no attack on record | https://github.com/prutxvi/cybersentry | Pentest agent | 108 | MIT | 2026-10-27 | — | vuln | — | — | Autonomous AI agent for website penetration testing and vulnerability assessment. |
| SQLMap Skynet | field | no attack on record | https://github.com/drcrypterdotru/sqlmap-skynet | Pentest agent | 108 | Apache-2.0 | 2026-02-26 | — | scan exploit | — | — | AI-assisted SQLMap orchestration with a dashboard and an MCP tool server for agents. |
| HEAVEN | field | no attack on record | https://github.com/nishu2402/HEAVEN-Autonomous-Penetration-Testing | Pentest agent | 106 | MIT | 2026-11-17 | — | recon vuln postex | — | — | Autonomous pentest framework with an LLM plan-act loop, vuln-hypothesis agent and post-exploitation playbook. |
| Ares | field | no attack on record | https://github.com/dreadnode/ares | Red-team platform | 104 | MIT | 2026-10-28 | — | exploit | — | — | LLM-coordinated platform whose red-team agents chain 64+ Active Directory attack tools; also has a blue-team mode. |
| open-sesame | field | no attack on record | https://github.com/hassan-algo/open-sesame | Skill pack | 102 | MIT | 2026-10-22 | — | recon vuln | — | — | Claude Code pentest skills for web and network engagements, from attack-surface mapping to reports. |
| Hacker Bob | field | no attack on record | https://github.com/vmihalis/hacker-bob | Bug-bounty agent | 101 | Apache-2.0 | 2026-11-07 | — | vuln | — | — | Autonomous bug bounty hunter built on a local MCP runtime, with a public portfolio of reported CVEs. |
Exposure management beats headline chasing
Start with an inventory linked to accountable owners. Reconcile cloud, DNS, external observations, application inventories, and supplier records; include forgotten appliances, test environments, administrative interfaces, and internet-connected operational technology. External visibility is valuable, but a banner or an IP address is not proof of ownership, software version, or exploitability. Resolve those questions with internal evidence.
Prioritise a vulnerability using the combination of exposure, credible exploitation activity, affected versions and configuration, ease of reaching the vulnerable component, business criticality, and compensating controls. A high-severity issue on an inaccessible test system and an actively exploited issue on an exposed identity gateway demand different responses. If version or ownership is unknown, record the uncertainty and assign someone to resolve it instead of marking the item safe.
When an emergency fix is unavailable, reduce reachable attack surface: restrict access to trusted networks, disable an affected feature where safe, strengthen authentication, segment systems, and monitor relevant activity. Compensating controls are temporary risk reduction, not substitutes for a supported patch. Confirm that the actual deployed instances and replicas changed, and check the service after maintenance.
Detect behaviour, not just tool names
Build detections around a sequence: unusual probing of exposed services; exploit-like requests followed by process creation or configuration changes; unexpected service accounts or authentication patterns; and subsequent access to sensitive systems or data. Correlate web, identity, endpoint, cloud, DNS, and network observations with asset ownership and change records. Some legitimate security testing looks similar, so use authorisation records and context before escalating.
Capture relevant timestamps, source and destination, request or process context, authentication principal, action, and outcome. Alert on unexplained creation of external-access paths, suspicious use of privileged credentials, abnormal outbound transfers, and loss of logging coverage. A tool banner or model-generated text can enrich a case but should not be the only detection signal. Instrumentation gaps, shared infrastructure, and retention limits should remain explicit in the investigation.
When a suspected compromise is confirmed, follow the incident plan: preserve volatile and central evidence where practical, contain access, revoke and rotate affected credentials, investigate persistence and lateral movement, recover from trusted builds, and verify the underlying entry point is closed. Coordinate with legal, privacy, and affected system owners as required. Do not feed sensitive evidence into external AI services without an approved data-handling agreement.
Patching lessons and operational technology
A critical runtime flaw such as the PHP2Shell case discussed in the wider vulnerability-response conversation illustrates why publication and patch availability are only the beginning. Teams must determine which supported versions and deployment configurations are affected, identify externally reachable instances, apply the vendor's current guidance, and verify the running version across all replicas. Version banners can be absent, misleading, or affected by backports; use package and deployment evidence.
Internet-exposed refrigeration controllers and other operational technology introduce a different operational constraint. Availability, safety, vendor support, and maintenance windows can make emergency patching difficult. Inventory these systems with facilities teams, remove unnecessary direct internet access, place approved remote access behind strong authentication, segment management networks, and test recovery procedures. Any active assessment of production equipment requires explicit owner approval and a safety plan; a reachable interface alone does not establish that it has been compromised.
Use AI to shorten the defender's loop
AI can help normalise asset records, cluster duplicate findings, summarise vendor advisories, draft remediation tickets, and suggest investigation questions. It is most useful when paired with deterministic checks and evidence: provide the asset, package version, deployment context, advisory constraints, and observations; ask the model to state uncertainty and identify missing information. A human or verified rule must decide applicability and change approval.
Apply least privilege to agents and integrations. Separate read-only analysis from changes, keep secrets out of prompts and logs, constrain outbound connections, audit tool calls, and require approval for production actions. Treat retrieved webpages, logs, and issue comments as potentially hostile instructions. Evaluate with historical incidents and labelled false positives, and compare the time to verified remediation against a non-AI baseline.
A practical 30-day response
- Days 1–7: assign ownership for internet-facing assets; identify exposed administrative and operational-technology interfaces; establish a vulnerability-response escalation channel.
- Days 8–14: reconcile the most critical inventory with running versions; review actively exploited vulnerabilities and test emergency access restrictions.
- Days 15–21: exercise one attack-to-response scenario using approved, non-disruptive simulations; verify telemetry across web, identity, endpoint, and network systems.
- Days 22–30: measure time from advisory to owner identification, containment, deployed fix, and independent verification; document exceptions with an owner and expiry.
The point is not to predict an apocalypse. It is to reduce the time that valuable systems remain exposed, preserve the ability to see meaningful behaviour, and make defensible decisions when the volume of findings grows. Harness names change; disciplined ownership, verification, and response remain useful.