AI Security

AI and MCP Security Assessment

PSC assesses LLM applications, AI agents, and Model Context Protocol integrations as connected systems rather than isolated models. The engagement maps prompts, tools, resources, credentials, memory, and approval boundaries, then tests whether untrusted content or a compromised connector can trigger unintended actions.

This assessment is for teams deploying copilots, autonomous workflows, MCP servers or clients, and tool-enabled agents. The outcome is a prioritized account of exploitable trust failures and practical controls for authorization, consent, isolation, monitoring, and human approval.

Subject-matter owner: Viral Maniar · Managing Director and Principal ConsultantReviewer: Shabnam Sharma · Product Owner and DevSecOps Advisor
Risk focus

Problems this engagement addresses

The scope follows credible attack and failure paths rather than a generic control checklist.

  • Direct and indirect prompt injection crossing application or agent trust boundaries.
  • MCP tools, resources, or descriptions that can be poisoned, replaced, or over-trusted.
  • Over-broad credentials, scopes, and consent paths that turn model output into privileged action.
  • Sensitive data leaking through prompts, retrieval, memory, logs, or tool responses.
  • Missing sandbox, rate, step, and human-approval controls around consequential operations.
Engagement fit

When this service is useful

A clear fit decision protects assessment quality and prevents a broad request from becoming a shallow exercise.

Designed for

  • Teams preparing an LLM feature or agent for production.
  • Organizations connecting models to internal data, APIs, code execution, or business systems.
  • MCP server and client owners who need authorization and connector-level assurance.
  • Security teams validating guardrails, monitoring, and incident evidence for AI workflows.

Not the right fit for

  • A model-quality or accuracy benchmark without a security objective.
  • A compliance certification or guarantee that all future model behavior is safe.
  • Unbounded testing against third-party models or services without written authorization.
Scope

Choose the boundary that matches the decision

Options can be combined after owners, environments, authority, safety limits, and expected evidence are agreed.

LLM application

Prompts, retrieval, output handling, identity, tenancy, data flows, and abuse controls around a chatbot or copilot.

Agent and tools

Planning loops, memory, delegated tools, credentials, side effects, approval gates, and failure containment.

MCP ecosystem

Servers, clients, transports, tools, resources, prompts, scopes, consent, and malicious-connector scenarios.

Prerequisites

  • An architecture and data-flow walkthrough with owners of the model, application, and connected tools.
  • Authorized test identities, representative roles, and a safe environment or agreed production safeguards.
  • Current system prompts, tool definitions, MCP configuration, and relevant logging access.
  • Written rules of engagement for data handling, external providers, and consequential actions.

Explicit exclusions

  • Training-data provenance or model-supply-chain review unless explicitly scoped.
  • Destructive, irreversible, or financially consequential tool actions.
  • Third-party infrastructure outside the client's authority.
  • Continuous monitoring after the agreed assessment window.
Methodology

Evidence at every phase

Each phase produces an observable input to the final risk and remediation decisions.

  1. Architecture and trust mapping

    Inventory models, prompts, retrieval sources, agents, MCP connectors, tools, credentials, memory, and human decision points.

  2. Threat modelling

    Define abuse cases for injection, goal hijack, confused-deputy behavior, excessive agency, data disclosure, poisoning, and denial of wallet.

  3. Adversarial testing

    Exercise direct and indirect injection, unsafe output handling, tool chaining, scope bypass, malicious resources, and connector replacement within the rules of engagement.

  4. Control validation

    Test authorization, consent, isolation, allow-lists, approval gates, rate and step limits, logging, and incident reconstruction.

  5. Reporting and retest

    Provide reproducible evidence, system-level remediation, residual-risk decisions, and verification of agreed fixes.

Standards and frameworks used

OWASP Top 10 for LLM ApplicationsOWASP Agentic AIMITRE ATLASNIST AI RMFMCP security guidance
Outputs

Deliverables built for decisions and fixes

Reporting separates observed evidence, uncertainty, business priority, and the action needed to verify remediation.

  • AI and MCP attack-surface inventory with trust boundaries.
  • Risk-ranked findings with reproducible prompts, connector conditions, and observed effects.
  • Tool and credential authorization matrix.
  • Guardrail, isolation, consent, logging, and human-approval recommendations.
  • Executive readout, technical workshop, and agreed remediation retest.

Severity approach

Severity combines exploitability, required access, action autonomy, data sensitivity, blast radius, reversibility, and the strength of compensating human or technical controls. Model output alone is not treated as impact unless it crosses a trust boundary or enables a credible harmful action.

Readout

PSC presents an executive risk narrative and a technical walkthrough for product, engineering, and security teams. Findings identify the affected trust boundary, required preconditions, evidence, and a testable remediation outcome.

Remediation support

Remediation support focuses on least-privilege tool design, explicit consent, output mediation, isolation, secret handling, observable action trails, and safe failure behavior. Recommendations distinguish application controls from provider-dependent limitations.

Retest policy

An agreed retest verifies corrected authorization, injection resistance, isolation, and approval behavior using the original evidence path. Material architecture changes are re-scoped rather than assumed equivalent.

Typical process

A governed path from scope to verification

Timing depends on scope, access, environment, and evidence quality; PSC confirms milestones during scoping rather than publishing unsupported duration promises.

01

Confirm objectives and authority

02

Map the connected AI system

03

Execute controlled abuse cases

04

Review evidence with builders

05

Verify prioritized fixes

FAQ

Questions buyers ask before scoping

Final rules, access, and deliverables are confirmed in the engagement scope.

Is an MCP assessment different from an API penetration test?

Yes. API testing remains relevant, but MCP adds model-mediated tool selection, prompt and resource content, consent, connector trust, and delegated authority. The assessment follows those additional paths from untrusted content to tool effects.

Can PSC test a production agent?

Only with explicit safeguards and rules of engagement. Consequential actions are normally redirected to a controlled environment, test tenant, mock tool, or approval gate so evidence can be collected without creating business impact.

Does the assessment guarantee that prompt injection is impossible?

No. The goal is to reduce reachable impact through layered controls, constrained authority, isolation, monitoring, and safe human decisions. Residual model uncertainty is documented rather than hidden.

What should we provide before testing?

Provide architecture and data flows, model and agent configuration, tool and MCP inventories, test roles, logging access, and owners who can explain intended workflows and approve safe test cases.

Scope this service

Discuss ai and mcp security assessment with PSC.

Share the system, decision, and constraints. PSC will confirm whether this service fits before proposing a scope.

Request a scoped assessment