Prove What Your Agents Did
A log records what a system reported about itself. Evidence records what happened, in a form a third party can verify.
Key Takeaways:
- A log is not evidence. Sampled logs, unlinked traces, and records held solely by the team that operates the agents will not survive an audit, an incident review, or opposing counsel.
- The regulation already applies. EU AI Act Article 12 has covered high-risk systems since August 2, 2026, with retention floors, explanation rights, and penalty tiers attached to records many organizations are not generating.
- Prompt injection hides in what the agent reads, not what it’s told. If you only log the user prompt and the model response, the attack never appears in your record at all.
- Hash chaining proves modification, not authorship. It detects a changed record after the fact. Key custody and independent verification are what make that detection mean something to an auditor, and that includes asking your own vendor how both work.
- A receipt beats a log. A structured record signed and chained at every step, capturing full input through output across every agent handoff, turns reconstruction from a multi-week effort into minutes.
Six months from now, someone asks what one of your agents did. Could you hand over a record and prove it is complete, prove it is unaltered, and prove both without asking anyone to take your word for it?
Most enterprises can offer logs, summaries, and a reasonable guess at what probably happened. Very few can do all three at once. The gap is architectural. Logging configurations sample under load, traces stop at the first hop when an agent delegates to a sub-agent, and when the team running the agents also administers the log store, the evidentiary value of what they hand over drops sharply.
Download the report for how a signed, hash-chained receipt trail closes the gap, a worked example tracing an indirect prompt injection attack through a four-agent chain, a self-audit against the questions an investigator will actually ask, and the specific EU AI Act, HIPAA, and SR 11-7 provisions your evidence needs to satisfy.