Injected instructions
A message talks the agent into it.
Agent Action Security Review
We try to break your AI agent before someone else does.
For teams launching an AI agent that can refund, change records or grant access. We attack one action and show you what got through, with the evidence.
For teams launching an AI agent that can refund, change records or grant access. We attack one action and show you what got through.
What we try to break
Four ways we try to break it.
Permissions that are too wide
The account can do more than it needs.
A skipped approval
The change goes through without sign-off.
A change with no record
Nothing is left for anyone to check.
Examples of what we try, not findings.
What you get
You keep the report and the evidence.
Findings with evidence
Each claim sits beside its proof.
Unknowns in writing
What we couldn't establish is named.
What to fix first
Prioritised recommendations.
A report you keep
Written report and readout. No login.
How it works
Three steps. Nothing happens without your sign-off.
- 01
Tell us the action
One action and the systems it touches. No secrets.
- 02
Agree the boundary
Fixed fee and scope in a signed SOW. We don't test until you give written authority.
- 03
Get the findings
Report and readout. If we don't get through, that's a result too.
Not a full pentest, certification, fixing or monitoring. See the full scope
Request the review
Start with a fit check. Don't send secrets.
Tell us one action, which systems it touches, and when you need to decide.
Discuss a review: engage@mail.witnessops.com