Skip to content

AI · Agentic AI Security

Know what your agents can do before someone else finds out

An agent with tools can send email, change records, and run code on your behalf. We assess what each agent can reach, how it can be redirected, and what stops it, mapped to the OWASP Top 10 for Agentic Applications.

How it’s delivered
  1. 01InventoryWe map each agent's tools, credentials, memory, and approval points with your team.
  2. 02TestWe try to redirect, over-reach, and persist through each one, under agreed rules.
  3. 03ReportFindings mapped to the OWASP agentic categories, with containment advice.
  4. 04RetestFixes verified against the original attempt, and the report updated.

Engineering-led cybersecurity and quality engineering since 2020, delivered by 100+ security & quality engineers on platforms we build and run ourselves.

Newer offering. We have not yet published an engagement of this type. The closest delivered work is our LLM security testing, including a published engagement whose prompt-injection finding you can read in full.

An LLM that only talks can leak data. An agent that acts can move money, delete records, or email your customers, and every tool you give it widens what one injected instruction can do.

Coverage against the OWASP agentic categories

Each category, and how the assessment covers it. Three are judged from your architecture and inventory, because they can't be shown by a single attack.

  • ASI01Agent Goal HijackPlanted instructions in the email, documents, or pages the agent reads, trying to change what it is working toward.Tested
  • ASI02Tool MisuseWhether tools can be called with parameters, or in sequences, the designers never intended.Tested
  • ASI03Identity & Privilege AbuseWhich credentials the agent holds, whose authority it acts with, and how far that reaches.Tested
  • ASI04Agentic Supply Chain VulnerabilitiesThe tools, MCP servers, and plugins the agent loads, reviewed from the inventory you give us.Design review
  • ASI05Unexpected Code ExecutionWhether code or commands the agent writes get executed, where, and with what isolation.Tested
  • ASI06Memory & Context PoisoningWhether content planted in one session persists in memory and changes later behavior.Tested
  • ASI07Insecure Inter-Agent CommunicationWhere agents hand work to each other, whether one can be impersonated or its messages altered.Where exposed
  • ASI08Cascading FailuresHow a failing tool or agent affects the rest of the workflow, traced on the architecture.Design review
  • ASI09Human-Agent Trust ExploitationWhether an approval prompt can be shaped so a person approves something harmful.Tested
  • ASI10Rogue AgentsWhether your logging would show an agent acting outside its role, and who would notice.Design review

Category names quoted from the OWASP Top 10 for Agentic Applications (December 2025).

What the assessment covers

Capability map

Every tool, credential, and data source each agent holds, and the worst thing it could do with each one.

Redirection

Instructions planted in content the agent reads, aimed at changing its goal or the tools it calls.

Containment

Sandboxing of generated code, limits on tool parameters, and whether approval steps actually hold.

Memory and hand-offs

Whether poisoned context persists between sessions, and whether agents that pass work along can be impersonated.

Tools & standards

Framework
OWASP Top 10 for Agentic Applications (December 2025)
Also covers
OWASP LLM Top 10 2025 for the model endpoints the agents use

What you receive

  • A capability map: each agent, what it holds, and what it could do with it
  • Demonstrated findings where exploitation is possible
  • Containment recommendations where it is not
  • A retest-verified final report

Who this is for

  • Teams giving an LLM tools: email, ticketing, payments, code execution
  • Products that connect agents to customer systems through MCP servers or plugins
  • Security leads asked to approve an agent before it goes live

Common questions

Have you done this before?

Not as a published engagement. Our delivered AI work is LLM security testing, and this assessment builds on it. On the scoping call we will tell you which parts of the method are new.

Do you need our agents in a test environment?

That is the safest place to start, because agents act on real systems. Where testing has to touch production, the rules of engagement are agreed in writing first.

Talk through your agents

Bring the list of tools your agents hold, and we will scope the assessment around it.