The engagement

The AI agent security assessment

Four to six weeks from the first workshop to verified fixes. We map what your agents can reach, attack them the way a motivated adversary would, implement the remediation with your engineers, then re-run the same attacks to prove they fail.

Typical duration

4-6 weeks

Scope unit

Per agent or agent fleet

Environment

Staging replica, or production under agreed rules

Output

Findings, fixes, verified re-test

Scope

Four domains, tested against your agents

The same four questions structure the whole engagement, from the inventory workshop to the final re-test.

What is the agent allowed to touch?

Permissions and identity

Agents inherit credentials from whoever wired them up, and that is usually a service account with far more reach than the task needs. We map the real blast radius of every agent identity.

  • What we test
  • Service accounts, OAuth scopes and API keys held by each agent
  • Standing access versus just-in-time, per-task credentials
  • Permission inheritance from the invoking user and from sub-agents
  • Secrets living in prompts, config files and tool definitions
  • What a single compromised agent could reach on its worst day
Phases

How the six weeks run

Nothing in this sequence waits on a tool finishing a scan. Each phase ends with something you can act on.

  1. 01Days 1 to 3

    Scope and inventory

    We build the map nobody has yet: every agent, the tools it holds, the data it reaches and the identity it runs as.

    • Workshop with the teams who built and operate the agents
    • Architecture, prompt, tool and credential review
    • A threat model tied to your business rather than a generic checklist
  2. 02Weeks 1 to 3

    Adversarial assessment

    We attack the agent end to end across the four domains: permissions, tool use, prompt injection and data egress.

    • Hands-on red teaming against a staging replica, or production under agreed rules
    • A tailored injection corpus built from your own untrusted inputs
    • Every finding reproducible, with the exact transcript that produced it
  3. 03Week 3

    Findings and risk ranking

    Findings ranked by what an attacker actually gains rather than by raw severity labels, with a briefing for executives and an appendix for engineers.

    • Severity tied to real blast radius and reachability
    • Mapped to the OWASP Top 10 for LLM Applications, MITRE ATLAS and NIST AI RMF
    • A remediation plan sequenced by effort against risk removed
  4. 04Weeks 3 to 6

    Remediation

    We implement the fixes alongside your engineers: scope reduction, tool guardrails, approval gates, injection defenses and egress control.

    • Pull requests and configuration changes against your own repositories
    • Least-privilege identities and just-in-time credentials for agents
    • Human-in-the-loop gates placed where they buy the most safety
  5. 05Week 6

    Verify and hand over

    We re-run the same attack suite against the fixed system and show, finding by finding, what no longer works.

    • A re-test report you can hand to a customer, an auditor or your board
    • The attack suite handed over as a regression pack you keep
    • A short enablement session for the engineering team
  6. 06Ongoing

    Continuous assurance

    Agents change every week. The suite runs on every release, and inventory drift such as new tools, new scopes or new MCP servers gets flagged.

    • Injection and permission regression tests in your CI pipeline
    • Drift alerts when an agent gains reach it did not have last week
    • A quarterly red-team refresh as new attack classes appear
Remediation

How we close the findings

Remediation happens alongside your engineers, in your repositories, through your review process. Most engagements involve some mix of the work below.

Least privilege & identity

  • A distinct, scoped identity per agent instead of a shared service account
  • Just-in-time, task-scoped credentials in place of standing access
  • Secrets moved out of prompts, tool definitions and config
  • Permission budgets that fail closed when an agent reaches past them

Tool & action guardrails

  • Tool allowlists with enforced argument schemas and typed validation
  • Human approval gates on irreversible, financial and outbound actions
  • Dry-run and idempotency support for high-impact tools
  • Depth, recursion and spend limits on chaining and sub-agents

Prompt-injection defenses

  • Hard trust boundaries between instructions and retrieved content
  • Provenance tagging and spotlighting for everything untrusted
  • Policy enforced at the tool-call layer, not only in the system prompt
  • Untrusted content handled by isolated agents that hold no privileged tools

Data & egress control

  • Retrieval that enforces the source system's access control per user
  • Tenant and session isolation across context, cache and memory
  • Redaction and secret detection on outputs, logs and traces
  • Domain allowlists on outbound calls, links and rendered content

Some fixes are product decisions rather than code changes: removing a capability, adding a confirmation step, narrowing who can invoke an agent. We bring you the trade-off with the evidence, and you decide.

Deliverables

What you get

Everything is reproducible, and everything is yours to keep.

Agent and tool inventory map

Every agent, tool, MCP server, data source and identity in one diagram. For most teams it is the first time anyone has seen the whole thing.

Ranked findings with reproductions

Each finding carries the transcript, the payload and the exact conditions that produce it. Nothing we cannot reproduce goes in the report.

Executive summary

A plain-English read of what an attacker could do today, written for the people who sign off on shipping the agent.

Implemented fixes

Pull requests, policies and configuration delivered straight into your repositories, so your team is not left with a to-do list.

Verification report

The same attacks, re-run after remediation, with a clear before and after for every finding.

Regression attack suite

The tests we built for you, handed over to run in CI so the fixes stay fixed.

Questions

What teams ask before signing

Something not covered here? Ask us directly.

Agent stacks we test

OpenAI Agents SDKAnthropic Claude and MCPLangChain / LangGraphCrewAIMicrosoft Copilot StudioAmazon Bedrock AgentsGoogle Vertex AI AgentsIn-house orchestration

Start with a scoping call

Thirty minutes on your agent architecture. You leave with our read on where the exposure sits and whether an assessment is worth the spend. No obligation either way.