Press enter or click to view image in full size
Detecting rogue AI agents is getting real, and the logs in your SIEM won’t come to the rescue.
Logs tell you what individual systems recorded. An agent’s actions cross authorization services, gateways, applications, and data stores. To understand one execution, you need to connect what each system decided, what it observed, and what happened next. That is difficult when records have different formats, missing context, and no reliable way to link them.
Evidence offers a more formalized runtime stream. Each producer makes a specific assertion about a decision or action, identifies the execution it belongs to, and signs the record. The record carries enough context to interpret it later. A collector can correlate independent assertions without claiming that any one producer saw the whole event.
This stream serves several purposes. It improves observability by showing how activity unfolded across systems. It supports accountability by identifying who made each claim. It gives risk management a way to spot concerning patterns and prioritize attention. It helps compliance teams verify that controls operated during actual executions.
It also gives AI better material for threat detection. An agent might explain away an unusual request or alter its own records. It cannot easily change what every downstream service independently recorded. Conflicting claims, missing steps, and unexpected outcomes become signals worth investigating.
We need to detect rogue agents. The Apple TV series “Murderbot” (based on the books by Martha Wells) is amazing at covering his hacker-tracks by deleting logs. By promoting the role of evidence in enterprise IT, we can make that fiction.