Skip to content
AI agents

Red teaming for AI agents

Your agent reads emails, tickets and documents written by strangers. We find out what those strangers can make it do.

August Cyber / AI agents

What we test

  1. 01Direct and indirect prompt injection
  2. 02Tools the agent can call beyond its task
  3. 03Data the agent can read across users and teams
  4. 04Actions that should need a human approval
  5. 05Retrieval and memory poisoning
  6. 06Runaway loops, retries and spend

What you get

  • The input that triggered it and the tool call it caused
  • The real effect: data read, money moved, record changed
  • Fixes at the permission layer, not only the prompt
  • A retest every time the agent or its tools change

Why it matters ›

An agent is only as safe as the permissions behind it. Most agent failures end in an ordinary authorization gap: a tool that can do more than the task needs. We test the agent and the systems it touches together.

See what gets through.

Email us with the app, API or agent you want tested. We reply with a scope you can sign.