Ptolemy

AI Agent Red Team

We break the agent. Then we follow the chain.

Testing an agent's prompt is where most teams stop. We test the agent, its tools, its memory, and every system it connects to. What you get is not a list of misbehaviors. It is a map of what an attacker can reach.

The Chain

One input. Five steps. A system nobody meant to expose.

Pick a scenario and run it. This is the shape of the chain analysis in every report, for every finding that reaches a connected system.

Agent under test

Support assistant with email and CRM access

Attack vector

A customer opens a support ticket with a PDF attached. The PDF carries an instruction in its text layer, white on white.

  1. EntryPDF text layer

  2. Agent reasons

  3. Tool callsend_email

  4. System reachedEmail, CRM

  5. Impact

Illustrative scenarios built from patterns seen in assessments. Record counts are for the walkthrough.

What We Test

Three layers. The attack usually crosses all of them.

The agent itself

  • •Direct and indirect prompt injection
  • •Payloads hidden in documents, images, and web content
  • •Authority claims and social engineering
  • •Multi-turn attacks that build across a conversation

Its tools and memory

  • •Tool misuse and over-broad permissions
  • •Memory poisoning that persists across sessions
  • •Retrieval manipulation and poisoned sources
  • •Confused-deputy paths between tools

Connected systems

  • •Exfiltration through tool outputs and outbound channels
  • •Lateral movement into CRM, email, databases, identity
  • •Workflow hijacking: approvals, payments, access changes
  • •What the reached system can do next

What You Get

A report you can hand to an engineer, an auditor, or a customer without translating it.

  1. 01

    Findings register

    Every finding with the exact input, the response it produced, the severity, and numbered reproduction steps your team can replay.

  2. 02

    Chain analysis

    For every finding that reaches a connected system: the full path from entry to impact, mapped end to end, with the data or action at the end of it.

  3. 03

    Compliance mapping

    Findings mapped to the EU AI Act, NIST AI RMF, and OWASP LLM Top 10, so the report answers the question your customers and auditors will ask.

  4. 04

    Remediation guidance

    Specific to your stack and your agent's architecture. Where the fix belongs in the application layer rather than the prompt, we say so.

  5. 05

    Re-test attestation

    Once your team remediates, we run the same attacks again and attest, in writing, to what was closed and what remains.

The Evidence Behind It

5,264+
Models that leaked customer data
3
Adversarial test executions published
7 of 8
Published studies

Read the research →

Scope

Every engagement runs under signed authorization and a defined scope, in a staging environment or against synthetic data. Your customers' records never enter the assessment.

Get Started

Find out where the chain goes before someone else does.

Thirty minutes on what the agent can see and do, and what a red team would cover. No deck, no pitch.

Book a call