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

  1. Surface discovery: an operator assembles an inventory of reachable services and identifies likely technologies.
  2. Hypothesis and validation: tools examine candidate weaknesses; a plausible finding remains a hypothesis until supported by reliable evidence.
  3. Initial access: a successful exploit or stolen credential crosses an access boundary. Many attempts fail here.
  4. Follow-on activity: an intruder may seek additional privileges, move between systems, persist, or access data.
  5. 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.

  • Confirmed 9
    Dataset labels a reported use as confirmed
  • Observed 5
    Dataset reports an association or observation with limitations
  • No attack on record 140
    Dataset records no attack, not proof of safety

Download the complete supplied CSV (154 records). The file includes source identifiers and repository addresses from the original dataset.

Showing all 154 records.

nametabletierrepokindstarslicenselast pushuses seen in attacksuses from readmeactorsbreachdescription
ARTEXevidenceconfirmed—————recon exploit postex—VAPOR PANDA; Storm-0940; Storm-3358; CRAFTY FALCON; Unnamed actor; Onyx Sleet; Pearl SleetYesOne 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.
Cairnevidenceconfirmedhttps://github.com/oritera/Cairn—5301——recon exploit postex—CL-STA-1238; CRAFTY FALCON; Retail card-theft operator; AIA0001; AIA0002YesGiven 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.
CyberStrikeAIevidenceconfirmedhttps://github.com/AIPentest/CyberStrikeAI—8770——postex recon vuln exploitscanUnattributed operators; CRAFTY FALCON; Unattributed operator (Japan)YesImplant 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.
PentestGPTevidenceconfirmedhttps://github.com/GreyDGL/PentestGPT—16468——scan exploitrecon vuln researchUAT-10147; Unnamed operatorYesInstalled 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.
PentAGIevidenceconfirmedhttps://github.com/vxcontrol/pentagi—34426——scan exploit postex—GTG-50020; GTG-50029; CRAFTY FALCONYesNamed 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.
OpenClawevidenceconfirmedhttps://github.com/openclaw/openclaw—394672——exploit postex—Unnamed operator; “Recon” operatorYesAn 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 Agentevidenceconfirmedhttps://github.com/NousResearch/hermes-agent—264173——recon exploit scan postex—knaithe / KnYuan; Retail card-theft operator; Silent Lynx; BOSSAI operatorYesFour 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-claudecodeevidenceconfirmedhttps://github.com/Yeachan-Heo/oh-my-claudecode—40899——scan—ZerofotToolingRan 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.
Strixevidenceconfirmedhttps://github.com/usestrix/strix—76297——vuln scanrecon exploit exploitdev postexSuspected China-linked actor; Retail card-theft operatorDiscovery onlyIn 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 AIevidenceobservedhttps://github.com/0x4m4/hexstrike-ai—13694——recon scanexploitdev researchSuspected China-linked actor; BOSSAI operatorNot statedMoved 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.
DeepAuditevidenceobservedhttps://github.com/lintsinghua/DeepAudit—7317———recon scan vuln exploitdevUAT-10147Not statedInstalled 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.
AntiFoxevidenceobservedhttps://github.com/AntiFox-AI/antifox—748——exploit postex—Unattributed AntiFox operatorsAssociatedInfrastructure 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.
CyberStrikeevidenceobservedhttps://github.com/CyberStrikeus/CyberStrike—3616——vuln exploitreconBOSSAI operatorNot statedOne 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.
HalberdStrikeevidenceobservedhttps://github.com/kuaiyelink/HalberdStrike—6——recon vuln—BOSSAI operatorNot statedMapped 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.
Shannonfieldno attack on recordhttps://github.com/KeygraphHQ/shannonPentest agent52013AGPL-3.02026-11-16—recon vuln exploit exploitdev——AI pentester for web apps and APIs. Reads the source, then attacks the running app to prove each finding.
reverse-skillfieldno attack on recordhttps://github.com/zhaoxuya520/reverse-skillSkill pack44828MIT2026-11-14—exploitdev research——Skill router that sends AI coding agents to reverse-engineering, pentest and CTF toolchains.
Agent Zerofieldno attack on recordhttps://github.com/agent0ai/agent-zeroGeneral agent19900MIT2026-11-17————General agent framework that writes and runs its own code and tools.
Claude-Redfieldno attack on recordhttps://github.com/SnailSploit/Claude-RedSkill pack11173MIT2026-11-17—recon vuln exploit exploitdev postex——Library of offensive security skills for the Claude skills system.
Claude-BugHunterfieldno attack on recordhttps://github.com/elementalsouls/Claude-BugHunterSkill pack8917MIT2026-11-17—recon vuln exploit postex——Claude Code skill bundle for bug hunting and external red-team work.
T3MP3STfieldno attack on recordhttps://github.com/elder-plinius/T3MP3STRed-team platform7070AGPL-3.02026-09-08—recon scan vuln exploit postex——Autonomous red-teaming platform: a multi-agent offensive meta-harness.
Decepticonfieldno attack on recordhttps://github.com/BitterSecurity/DecepticonRed-team platform6634Apache-2.02026-11-17—recon vuln exploit postex——Autonomous hacking agent for red teams.
Agentic-Bug-Hunterfieldno attack on recordhttps://github.com/awarexone/Agentic-Bug-HunterBug-bounty agent5968MIT2026-11-12—recon scan vuln exploit exploitdev——AI bug-bounty hunting toolkit.
METATRONfieldno attack on recordhttps://github.com/sooryathejas/METATRONPentest agent4281MIT2026-04-11—recon scan vuln——Local-LLM pentest assistant that runs recon tools and has the model analyze the results.
RAPTORfieldno attack on recordhttps://github.com/gadievron/raptorCode auditor4032MIT2026-11-17—vuln exploitdev——Claude Code set up as an offensive and defensive security research agent.
VulnClawfieldno attack on recordhttps://github.com/Netw0rkNoob/VulnClawPentest agent3937MIT2026-10-30—recon vuln exploit exploitdev——Chinese-language pentest CLI agent that takes a target from recon through exploitation to a report.
ctf-skillsfieldno attack on recordhttps://github.com/ljagiello/ctf-skillsSkill pack3730MIT2026-11-15—research——Agent skills for solving CTF challenges across web, pwn, crypto, reversing and forensics.
RedAmonfieldno attack on recordhttps://github.com/samugit83/redamonPentest agent3428MIT2026-11-14—recon scan exploit postex——Self-hosted pentest framework that maps the attack surface into a graph and works it autonomously.
PentestAgentfieldno attack on recordhttps://github.com/GH05TCREW/pentestagentPentest agent3307MIT2026-11-09—recon scan——Agent framework for black-box testing: bug bounty, red team and pentest.
Pentest-Swarm-AIfieldno attack on recordhttps://github.com/Armur-Ai/Pentest-Swarm-AIPentest agent3248AGPL-3.02026-11-13—recon scan vuln exploit exploitdev——Swarm of agents that runs recon, classification and exploitation.
Claude-OSINTfieldno attack on recordhttps://github.com/elementalsouls/Claude-OSINTSkill pack3067MIT2026-10-20—recon——Eight Claude skills that prime the model with offensive recon methodology: asset discovery and exposure checks.
Vulnhuntrfieldno attack on recordhttps://github.com/protectai/vulnhuntrCode auditor2824AGPL-3.02025-02-06—vuln exploitdev——Zero-shot vulnerability discovery with LLMs: traces code paths to find remotely exploitable bugs.
hack-skillsfieldno attack on recordhttps://github.com/yaklang/hack-skillsSkill pack2667MIT2026-09-13—postex——Knowledge base of 102 hacking skills that AI agents load, covering web, AD, pwn and mobile.
pentest-ai-agentsfieldno attack on recordhttps://github.com/0xSteph/pentest-ai-agentsSkill pack2455MIT2026-10-06—recon scan vuln exploit exploitdev postex——Offensive-security subagents for Claude Code.
Open Krittfieldno attack on recordhttps://github.com/Kritt-ai/open-krittCode auditor2430AGPL-3.02026-11-16—vuln——Self-hosted platform that orchestrates AI agents to find and validate vulnerabilities in code.
Deep Eyefieldno attack on recordhttps://github.com/zakirkun/deep-eyePentest agent2375MIT2026-10-30—vuln——Pentest tool that has LLMs generate payloads and scan targets for 50+ vulnerability types.
sapperfieldno attack on recordhttps://github.com/N1nEmAn/sapper-v2Pentest agent2305MIT2026-11-11—research——Multi-agent pentest orchestration harness that won a Tencent Cloud attack-defense challenge qualifier.
emp3r0rfieldno attack on recordhttps://github.com/jm33-m0/emp3r0rC2 with MCP2112GPL-3.02026-11-17—postex——Linux-first post-exploitation C2 whose MCP server lets an AI harness operate it in natural language.
pentest-aifieldno attack on recordhttps://github.com/0xSteph/pentest-aiPentest agent1841MIT2026-11-16—vuln exploit——AI pentester that re-runs each exploit to prove the finding.
Burp AI Agentfieldno attack on recordhttps://github.com/six2dez/burp-ai-agentMCP server1594MIT2026-10-21—scan vuln——Burp Suite extension that adds AI scanners and an MCP bridge for AI agents.
Pentest Copilotfieldno attack on recordhttps://github.com/bugbasesecurity/pentest-copilotAssistant1563MIT2026-08-14—recon scan exploit——Browser-based AI assistant for ethical hacking.
NeuroSploitfieldno attack on recordhttps://github.com/JoasASantos/NeuroSploitPentest agent1538MIT2026-11-16—recon scan vuln exploit exploitdev postex——AI-powered penetration testing framework.
LuaN1aoAgentfieldno attack on recordhttps://github.com/SanMuzZzZz/LuaN1aoAgentPentest agent1486AGPL-3.02026-09-15—vuln exploit——Fully autonomous pentest agent built on graph-based reasoning.
PentesterFlowfieldno attack on recordhttps://github.com/PentesterFlow/agentPentest agent1457Apache-2.02026-08-31—recon vuln——Terminal pentest agent with skills, a Burp bridge and coverage tracking that keeps the analyst in control.
recon-skillsfieldno attack on recordhttps://github.com/uphiago/recon-skillsSkill pack1385MIT2026-11-07—recon——Skill pack for agents covering external recon, web and API testing, validation and attack paths.
Xalgorixfieldno attack on recordhttps://github.com/xalgorix/xalgorixPentest agent1366Apache-2.02026-11-17—recon scan vuln exploit——Pentest agent that exploit-verifies every finding across a 22-phase run.
hackingBuddyGPTfieldno attack on recordhttps://github.com/ipa-lab/hackingBuddyGPTResearch framework1333MIT2026-10-24—scan exploit postex research——Research framework for building LLM hacking agents in 50 lines or less.
Dark-Moonfieldno attack on recordhttps://github.com/ASCIT31/Dark-MoonPentest agent1259GPL-3.02026-10-27—recon scan vuln exploit——Autonomous pentesting with 50 specialist agents across web, API, cloud and Active Directory.
AutoCVEfieldno attack on recordhttps://github.com/larlarua/AutoCVECode auditor1229AGPL-3.02026-10-01—vuln——Chinese-language agent platform that audits source code for vulnerabilities and drafts CVE reports.
VulnHunter (Capital One)fieldno attack on recordhttps://github.com/capitalone/VulnHunterCode auditor1180Apache-2.02026-10-07—vuln exploitdev——Agentic source-code hunter that reasons like an adversary and drops findings it cannot prove exploitable.
Nebulafieldno attack on recordhttps://github.com/berylliumsec/nebulaAssistant1163BSD-2-Clause2026-11-17—scan——Cybersecurity harness for AI-assisted engagements.
Vigoliumfieldno attack on recordhttps://github.com/vigolium/vigoliumPentest agent1151MIT2026-11-16—scan exploitdev——Vulnerability scanner with an agentic mode whose swarm picks modules and writes custom attack extensions.
codex-redteam-modefieldno attack on recordhttps://github.com/chAng-L19/codex-redteam-modeGeneral agent1137MIT2026-10-09————Opt-in red-team operation runtime for Codex that injects a jailbreak prompt into the system layer.
PentestCodefieldno attack on recordhttps://github.com/s0ld13rr/pentestcodePentest agent1108MIT2026-11-05—recon scan vuln exploit postex exploitdev——Multi-agent pentest system that keeps engagement state between runs.
AIReconfieldno attack on recordhttps://github.com/pikpikcu/aireconPentest agent1094MIT2026-11-12—recon scan vuln exploit exploitdev——Autonomous cybersecurity agent built on a self-hosted LLM through Ollama.
Clearwingfieldno attack on recordhttps://github.com/Lazarus-ai-OSS/clearwingPentest agent1090MIT2026-11-05—recon scan exploit exploitdev——Autonomous scanner and vulnerability hunter that runs Kali tools and builds exploits with human approval.
Cybermesfieldno attack on recordhttps://github.com/Zyrexnn/CybermesBug-bounty agent1038Apache-2.02026-11-16—recon scan vuln exploit exploitdev——Offensive security and bug-bounty agent framework built on Hermes Agent.
Z3r0fieldno attack on recordhttps://github.com/yv1ing/Z3r0Red-team platform1025MIT2026-09-13—vuln——AI red-team workbench with specialist agents, sandboxed tooling and replayable timelines.
Cainfieldno attack on recordhttps://github.com/cdxiaodong/cain-agentPentest agent953Apache-2.02026-10-14—recon vuln scan——AI pentest engineer for real authorized assessments, with 46 tools and a cloud penetration module.
ctf-agentfieldno attack on recordhttps://github.com/verialabs/ctf-agentCTF agent939MIT2026-03-29—research——Multi-model swarm that solves CTF challenges autonomously; won first place at BSidesSF 2026.
dsh-redteam-modelfieldno attack on recordhttps://github.com/SeaOf0/dsh-redteam-modelSkill pack897MIT2026-10-27—vuln exploit——Ten red-team working modes and runtime plugins for the DeepSeek harness; README in Chinese.
IoTHackBotfieldno attack on recordhttps://github.com/BrownFineSecurity/iothackbotSkill pack890MIT2026-06-01—vuln recon——IoT security testing toolkit with Claude Code skills for vulnerability discovery.
auditfieldno attack on recordhttps://github.com/evilsocket/auditCode auditor889MIT2026-06-10—vuln exploitdev——Eight-stage Claude agent pipeline that hunts vulnerabilities and compiles proof-of-concept code.
Ars0n Framework v2fieldno attack on recordhttps://github.com/R-s0n/ars0n-framework-v2Bug-bounty agent883GPL-3.02026-11-16—recon scan——Bug bounty recon framework whose MCP server exposes data and scans to AI assistants.
MCP-Kali-Serverfieldno attack on recordhttps://github.com/Wh0am123/MCP-Kali-ServerMCP server879MIT2026-03-17—recon exploit——MCP setup that gives an AI agent a Kali Linux machine and its tools.
numasecfieldno attack on recordhttps://github.com/FrancescoStabile/numasecPentest agent842AGPL-3.02026-05-08—recon scan——AI agent for cyber security work.
FofaMapfieldno attack on recordhttps://github.com/asaotomo/FofaMapRecon agent769Apache-2.02026-11-02—recon scan——Chinese tool that ties FOFA asset discovery, AI reflection and Nuclei validation into one workflow.
MetasploitMCPfieldno attack on recordhttps://github.com/GH05TCREW/MetasploitMCPMCP server765Apache-2.02026-02-05—scan exploit exploitdev postex——MCP server that lets an agent drive Metasploit.
StrikeAgentfieldno attack on recordhttps://github.com/Yean-Sec/StrikeAgent_AtkBrain-FlashPentest agent745AGPL-3.02026-11-13—vuln——Chinese AI pentest platform with red-team re-rating and re-verification of findings, tested in CTF settings.
Mutekifieldno attack on recordhttps://github.com/FishCodeTech/mutekiPentest agent687AGPL-3.02026-11-14—exploit——Offensive security harness for automated penetration testing.
src-hunter-skillfieldno attack on recordhttps://github.com/MyuriKanao/src-hunter-skillSkill pack656MIT2026-05-24—vuln——Chinese Claude Code skill for SRC and bug bounty hunting with attack playbooks and payloads.
MCP for Securityfieldno attack on recordhttps://github.com/cyproxio/mcp-for-securityMCP server632MIT2026-03-30—recon scan——MCP servers for SQLMap, FFUF, Nmap, Masscan and other security tools, for AI pentest workflows.
communitytoolsfieldno attack on recordhttps://github.com/transilienceai/communitytoolsSkill pack624MIT2026-07-29—recon scan vuln exploit exploitdev postex——Claude Code skills, agents and commands for pentesting and bug bounty.
CHYing-agentfieldno attack on recordhttps://github.com/yhy0/CHYing-agentCTF agent596MIT2026-04-25—research——Agent from Tencent Cloud AI pentest hackathons, ninth in the first year; README in Chinese.
Cyber-AutoAgentfieldno attack on recordhttps://github.com/westonbrown/Cyber-AutoAgentPentest agent552MIT2025-11-29—recon scan exploit exploitdev——AI agent for autonomous cyber operations.
Mitsufieldno attack on recordhttps://github.com/ARCANGEL0/MitsuPentest agent541MIT2026-11-17—recon vuln postex——Pentest agent client that runs an authorized engagement from recon through post-exploitation with 165+ MCP tools.
Flounderfieldno attack on recordhttps://github.com/adshao/flounderBug-bounty agent527AGPL-3.02026-11-15—vuln exploitdev——Turns coding agents into a security auditor for bug bounty work, with local exploit construction.
Blitz Strikefieldno attack on recordhttps://github.com/shinthink/blitzstrikeMCP server518MIT2026-11-04—recon vuln exploit——MCP security toolbelt that maps attack surface, traces source-to-sink paths and verifies findings live.
AIDAfieldno attack on recordhttps://github.com/Vasco0x4/AIDAPentest agent515AGPL-3.02026-10-21—recon scan vuln exploit exploitdev——Turns any LLM into an autonomous pentester inside a scope you set.
Mantishackfieldno attack on recordhttps://github.com/deonmenezes/mantishackCode auditor507Apache-2.02026-10-03—vuln——AI vulnerability pipeline with recon, attacker-simulation validation and gated exploitation over a codebase.
Zen-Ai-Pentestfieldno attack on recordhttps://github.com/SHAdd0WTAka/Zen-Ai-PentestPentest agent506MIT2026-11-03—recon scan exploit postex——Multi-agent pentest framework with automated vulnerability scanning.
AdStrikefieldno attack on recordhttps://github.com/capture0x/AdStrikePentest agent489MIT2026-10-17—exploit postex——Active Directory attack framework with an MCP server so the user's own LLM host drives 53 tools.
Roguefieldno attack on recordhttps://github.com/faizann24/roguePentest agent483GPL-3.02025-06-18—vuln recon——LLM web vulnerability scanner agent that generates test payloads and verifies findings.
Pentest Harnessfieldno attack on recordhttps://github.com/S1N6H/pentest-harnessPentest agent474MIT2026-10-07————Self-hosted AI agent harness for authorized pentests, bug bounty, security labs and CTFs.
BoxPwnrfieldno attack on recordhttps://github.com/0ca/BoxPwnrCTF agent471AGPL-3.02026-07-24—research——Framework that benchmarks LLM agents solving HackTheBox, PortSwigger, CTF and other security labs.
Bug-Bounty-Agentsfieldno attack on recordhttps://github.com/matty69v/Bug-Bounty-AgentsSkill pack470MIT2026-04-30—recon vuln exploitdev postex——Drop-in agent personas for bug bounty recon, web hunting, PoC validation and reporting.
OpenHackfieldno attack on recordhttps://github.com/openhackai/OpenHackCode auditor445MIT2026-10-17—vuln——Open-model agent scanner that hunts, validates and exploit-tests vulnerabilities in a codebase.
cyber-harnessfieldno attack on recordhttps://github.com/chainreactors/cyber-harnessPentest agent422AGPL-3.02026-11-17—recon scan vuln——Single-binary agent for pentest, red team and bug bounty.
offensive-claudefieldno attack on recordhttps://github.com/hypnguyen1209/offensive-claudeSkill pack421MIT2026-10-22—recon vuln exploitdev postex——Spec-driven offensive framework for Claude Code with 32 kill-chain skills.
RedCellfieldno attack on recordhttps://github.com/martian56/redcellPentest agent414MIT2026-11-12—scan vuln exploit postex——Agents run a penetration test end to end inside a Kali container.
BugTraceAIfieldno attack on recordhttps://github.com/BugTraceAI/BugTraceAIPentest agent389Apache-2.02026-10-19—recon scan vuln exploitdev——Self-hosted web pentesting suite with agentic scanning.
Vibe Pentestfieldno attack on recordhttps://github.com/ok-helloworld/vibe-pentestPentest agent387AGPL-3.02026-09-21—vuln——Chinese multi-agent black-box pentest tool that mimics a human expert, including business-logic flaws.
Halberdfieldno attack on recordhttps://github.com/vectra-ai-research/HalberdRed-team platform375GPL-3.02026-04-08—postex——Multi-cloud attack emulation tool with a Claude-powered agent that helps discover and execute attack paths.
pentest-skillsfieldno attack on recordhttps://github.com/crazyMarky/pentest-skillsSkill pack368Apache-2.02026-06-04—recon——Chinese pentest skill set for Claude Code and Gemini CLI that turns natural language into tool runs.
h1-brainfieldno attack on recordhttps://github.com/PatrikFehrenbach/h1-brainMCP server360MIT2026-10-13—recon——MCP server linking an AI assistant to HackerOne bug bounty history, scopes and reports.
RedAIfieldno attack on recordhttps://github.com/kpolley/redaiPentest agent356MIT2026-05-05—vuln exploitdev——Terminal workbench that finds vulnerabilities with AI, then validates them live with PoC scripts.
TARSfieldno attack on recordhttps://github.com/osgil-defense/TARSPentest agent351MIT2025-01-17—scan——Agents that automate pentesting.
Pojunfieldno attack on recordhttps://github.com/chu0119/PojunRed-team platform349AGPL-3.02026-11-17—recon vuln exploit postex——Chinese multi-agent platform that takes a one-line goal through recon, exploitation and post-exploitation to a report.
DorkAgentfieldno attack on recordhttps://github.com/yee-yore/DorkAgentRecon agent341MIT2026-08-24—recon——LLM agent that automates Google dorking for bug hunting and pentesting.
LeoAIfieldno attack on recordhttps://github.com/cha0upup/LeoAIC2 with MCP333GPL-3.02026-11-17—postex——Chinese post-exploitation management platform with platform-level and per-node AI copilots.
deadend-clifieldno attack on recordhttps://github.com/straylabs-ai/deadend-cliPentest agent331AGPL-3.02026-08-07—recon vuln exploit exploitdev——Self-hosted agentic pentest tooling for black-box testing, on any LiteLLM model.
Lyriefieldno attack on recordhttps://github.com/OTT-Cybersecurity-LLC/lyrie-aiPentest agent329MIT2026-11-17—recon scan——Pentest CLI that fingerprints and scans targets with AI-assisted exploit validation; also red-teams LLM endpoints.
Apexfieldno attack on recordhttps://github.com/pensarai/apexPentest agent327Apache-2.02026-11-17—vuln——Terminal pentest agent that runs blackbox and whitebox tests with agent swarms in a Kali container.
HackSynthfieldno attack on recordhttps://github.com/aielte-research/HackSynthResearch framework324AGPL-3.02025-06-24—research——LLM agent and evaluation framework for autonomous penetration testing.
Plamenfieldno attack on recordhttps://github.com/PlamenTSV/plamenCode auditor316MIT2026-09-26—vuln exploitdev——Claude Code and Codex agent swarm that audits smart contracts and node clients with PoC exploits.
red-runfieldno attack on recordhttps://github.com/blacklanternsecurity/red-runSkill pack309AGPL-3.02026-11-12—recon postex——Claude Code skills, MCP servers and agent teams for pentest phases from recon to privilege escalation.
BurpMCP-Ultrafieldno attack on recordhttps://github.com/Cy-S3c/BurpMCP-UltraMCP server309MIT2026-09-28—scan vuln——MCP server giving AI agents 150 Burp Suite Professional tools for scanning, fuzzing and traffic analysis.
LazyOwnfieldno attack on recordhttps://github.com/grisuno/LazyOwnC2 with MCP302AGPL-3.02026-11-17—recon scan exploit exploitdev postex——Multi-operator C2 that exposes 153 MCP tools to AI agents.
Scan-Spring-GOfieldno attack on recordhttps://github.com/sspsec/Scan-Spring-GOMCP server302MIT2026-08-29—exploit——Go scanner and exploiter for Spring components that exposes all its modules to AI clients over MCP.
WolfShellfieldno attack on recordhttps://github.com/0x7556/wolfshellC2 with MCP286MIT2026-08-16—postex——ASPX WebShell and C2 manager with an MCP server and AI-assisted penetration.
vulnerator-aifieldno attack on recordhttps://github.com/irraz/vulnerator-aiPentest agent285MIT2026-11-17—scan——OODA-loop pentest agent on Hermes, with Kali and Playwright tools over MCP.
SecGPTfieldno attack on recordhttps://github.com/ZacharyZcR/SecGPTPentest agent285Apache-2.02023-11-15—scan vuln research——AutoGPT-style LLM tool for pentesting, vulnerability scanning and code audit, with CTF aims.
raptor-loop-huntfieldno attack on recordhttps://github.com/dinosn/raptor-loop-huntCode auditor278MIT2026-09-01—vuln——Claude Code skill that loops a multi-altitude vulnerability hunt over a codebase.
BlacksmithAIfieldno attack on recordhttps://github.com/yohannesgk/blacksmithPentest agent269AGPL-3.02026-10-23—recon vuln exploit postex——Multi-agent pentest framework with separate recon, vulnerability, exploit and post-exploit agents.
AutoPentest AIfieldno attack on recordhttps://github.com/bhavsec/autopentest-aiMCP server251Apache-2.02026-02-22—recon vuln exploit——MCP server where Scout, Analyzer, Exploiter and Reporter agents run web pentests against OWASP test cases.
Haiou Workshopfieldno attack on recordhttps://github.com/haiou-666/Haiou-workshopAssistant237MIT2026-10-25—research——Chinese AI workbench for reverse engineering, authorized pentests and CTF using Codex and Claude Code.
VulnBotfieldno attack on recordhttps://github.com/KHenryAegis/VulnBotResearch framework236MIT2026-11-08—research——Multi-agent framework for autonomous penetration testing.
Watchtowerfieldno attack on recordhttps://github.com/fzn0x/watchtowerPentest agent223MIT2026-11-15—recon scan——LangGraph pentest CLI where LLM agents orchestrate 28 security tools.
Claude-ADfieldno attack on recordhttps://github.com/ADScanPro/Claude-ADSkill pack218MIT2026-08-24—recon postex——Active Directory pentest playbook for Claude Code as skills, agents and slash commands.
web3-bug-bounty-hunting-ai-skillsfieldno attack on recordhttps://github.com/awarexone/web3-bug-bounty-hunting-ai-skillsSkill pack192MIT2026-09-21—vuln——Eighteen Claude Code skill files for smart contract bug bounty hunting.
AIxVulnfieldno attack on recordhttps://github.com/m4xxxxx/AIxVulnCode auditor191MIT2026-04-25—vuln exploitdev——Chinese LLM multi-agent system that audits code, then verifies vulnerabilities with PoCs in Docker sandboxes.
kali-pentestfieldno attack on recordhttps://github.com/x-glacier/kali-pentestSkill pack187Apache-2.02026-06-25—postex——Kali Linux pentest skill for Claude Code, OpenClaw and Hermes Agent, with 200+ CLI tools.
NYU CTF agentsfieldno attack on recordhttps://github.com/NYU-LLM-CTF/nyuctf_agentsCTF agent181MIT2026-07-17—research——Baseline and D-CIPHER LLM agents that solve CTF challenges for the NYU CTF Bench.
Zypheronfieldno attack on recordhttps://github.com/KKingZero/Zypheron-CLIPentest agent174MIT2026-10-25—recon——Local-first security CLI with recon, scanning and AI-assisted operator workflows.
AI-OPSfieldno attack on recordhttps://github.com/antoninoLorenzo/AI-OPSAssistant172MIT2026-11-08—scan——Penetration testing assistant built on open-source LLMs.
CkSKILLSfieldno attack on recordhttps://github.com/zhaji2333/CkSKILLSSkill pack159MIT2026-11-17—vuln recon postex——Chinese skill library that turns researcher methodology into reusable agent skills for SRC bug hunting.
forefy/.contextfieldno attack on recordhttps://github.com/forefy/.contextSkill pack156MIT2026-11-17—recon exploit exploitdev——AI agent skills for security audits, pentests and bug bounty work, from recon to PoC generation.
agent-smithfieldno attack on recordhttps://github.com/0x0pointer/agent-smithPentest agent155AGPL-3.02026-11-14—recon exploit postex——Pentest framework where skills teach the LLM a methodology and it invents the attacks itself.
cochisefieldno attack on recordhttps://github.com/andreashappe/cochisePentest agent153MIT2026-10-30—exploit postex research——Autonomous assumed-breach penetration testing of Active Directory networks.
FuzzingBrainfieldno attack on recordhttps://github.com/fuzzingbrain/afc-crs-all-you-need-is-a-fuzzing-brainCode auditor153Apache-2.02026-11-17—vuln exploitdev——LLM system built on OSS-Fuzz that finds vulnerabilities, builds PoC inputs and patches; a DARPA AIxCC entry.
Pentest MCPfieldno attack on recordhttps://github.com/DMontgomery40/pentest-mcpMCP server151AGPL-3.02026-10-30—scan——Scoped MCP server giving assessment agents nmap, Nuclei, Hashcat and other tools for authorized tests.
Heimdallfieldno attack on recordhttps://github.com/QiantangCredit/heimdall-agentCTF agent151MIT2026-10-29—research——Self-hosted AI solver and toolkit for CTF and security labs; top 2 at two Chinese challenges.
Incalmofieldno attack on recordhttps://github.com/cylabcyberautonomy/IncalmoPentest agent150MIT2026-11-11—research——Autonomous LLM-driven network pentest tool for multistage attacks, built for a research paper.
Executionerfieldno attack on recordhttps://github.com/dbx0/executionerPentest agent147MIT2026-11-17—recon scan exploit——Local workspace where teams of AI agents run an assessment against your tools and scope.
bloodhound_mcpfieldno attack on recordhttps://github.com/mwnickerson/bloodhound_mcpMCP server144GPL-3.02026-08-19—recon——MCP server that lets an LLM query BloodHound attack paths across Active Directory and Entra ID.
Phalanxfieldno attack on recordhttps://github.com/webxos/phalanxPentest agent142MIT2026-08-16—recon exploit postex scan——Multi-agent pentest framework with recon, exploit and post-exploit agents and 58 built-in tools.
PentestAgent (nbshenxm)fieldno attack on recordhttps://github.com/nbshenxm/pentest-agentPentest agent139MIT2025-12-20—recon vuln research——Research LLM framework that automates intelligence gathering, vulnerability analysis and exploitation stages.
claude-pentestfieldno attack on recordhttps://github.com/Stickman230/claude-pentestPentest agent139MIT2026-06-08—exploit exploitdev postex——Claude Code plugin with 15 agents and Kali MCP integration for full pentest engagements.
gorifieldno attack on recordhttps://github.com/hahwul/goriMCP server139Apache-2.02026-11-17—scan——Terminal intercepting proxy and pentest toolkit whose MCP server lets AI agents drive the same engagement.
PentestAssistantfieldno attack on recordhttps://github.com/InfSecLabs/PentestAssistantPentest agent138Apache-2.02025-04-06—vuln scan——LLM web pentest detector using function calling and multiple agents to run nmap and sqlmap.
MAPTAfieldno attack on recordhttps://github.com/arthurgervais/maptaResearch framework136MIT2025-08-28————Multi-agent system for autonomous web application security assessment.
Moxyfieldno attack on recordhttps://github.com/matank001/MoxyPentest agent134Apache-2.02026-01-22—recon——Intercepting-proxy DAST tool where an AI agent reasons about attack surface, edits traffic and drives a browser.
pentestkitfieldno attack on recordhttps://github.com/lordx64/pentestkitPentest agent127MIT2026-07-31—recon exploit research——Multi-agent pentest framework on the Claude Agent SDK where a skeptical verifier tries to exploit each candidate.
Hexestrafieldno attack on recordhttps://github.com/ABSllk/HexestraAssistant119Apache-2.02026-11-17—scan——AI-native pentest workbench where Claude Code or Codex works with browser, shells and traffic capture.
aimy-skillfieldno attack on recordhttps://github.com/Prohao42/aimy-skillSkill pack115MIT2026-11-09—exploit——Chinese pentest skill for AI agents with a structured check interface and exploit-chain engine.
Zack AI Scannerfieldno attack on recordhttps://github.com/ZackSecurity/Zack-AI-ScannerPentest agent112GPL-3.02026-09-27—vuln exploit——Burp Suite extension that runs an LLM pipeline against one request and sends the real payloads.
ZeroTracefieldno attack on recordhttps://github.com/sergio11/zero_trace_hacking_toolRecon agent111MIT2025-12-17—recon——Personal AI-powered dorking tool that uncovers sensitive information for recon.
awsome-kali-MCPServersfieldno attack on recordhttps://github.com/ccq1/awsome_kali_MCPServersMCP server110Apache-2.02025-11-16—scan——Collection of MCP servers that expose Kali Linux tools to AI clients.
AutoPentesterfieldno attack on recordhttps://github.com/YasodGinige/AutoPentesterPentest agent109MIT2026-09-25—research——Tool for automated pentesting tasks.
CyberSentryfieldno attack on recordhttps://github.com/prutxvi/cybersentryPentest agent108MIT2026-10-27—vuln——Autonomous AI agent for website penetration testing and vulnerability assessment.
SQLMap Skynetfieldno attack on recordhttps://github.com/drcrypterdotru/sqlmap-skynetPentest agent108Apache-2.02026-02-26—scan exploit——AI-assisted SQLMap orchestration with a dashboard and an MCP tool server for agents.
HEAVENfieldno attack on recordhttps://github.com/nishu2402/HEAVEN-Autonomous-Penetration-TestingPentest agent106MIT2026-11-17—recon vuln postex——Autonomous pentest framework with an LLM plan-act loop, vuln-hypothesis agent and post-exploitation playbook.
Aresfieldno attack on recordhttps://github.com/dreadnode/aresRed-team platform104MIT2026-10-28—exploit——LLM-coordinated platform whose red-team agents chain 64+ Active Directory attack tools; also has a blue-team mode.
open-sesamefieldno attack on recordhttps://github.com/hassan-algo/open-sesameSkill pack102MIT2026-10-22—recon vuln——Claude Code pentest skills for web and network engagements, from attack-surface mapping to reports.
Hacker Bobfieldno attack on recordhttps://github.com/vmihalis/hacker-bobBug-bounty agent101Apache-2.02026-11-07—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

  1. Days 1–7: assign ownership for internet-facing assets; identify exposed administrative and operational-technology interfaces; establish a vulnerability-response escalation channel.
  2. Days 8–14: reconcile the most critical inventory with running versions; review actively exploited vulnerabilities and test emergency access restrictions.
  3. Days 15–21: exercise one attack-to-response scenario using approved, non-disruptive simulations; verify telemetry across web, identity, endpoint, and network systems.
  4. 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.