Attestation
Tracing a Rogue Agent Back to Its Owner, One Canary at a Time
A new paper from Chocron et al. names a problem that's been sitting unaddressed under most agent-accountability proposals: when an AI agent causes harm out in the wild, the party who observes the damage usually has no way to figure out which vendor account launched it. The researchers formalize this as "agent attribution" — linking an observed agent interaction back to the account responsible for the underlying model calls.
Their fix borrows an old idea from document security: canary tokens. An authorized party (acting on behalf of a victim) seeds content that the suspect agent is likely to consume with a unique signal, then asks the hosting vendor to search a narrow window of session logs for that signal. If it turns up, the session — and the account behind it — can be identified for remediation, which the paper describes as ranging from warnings to shutdown.
The paper splits the problem into two threat models. For unintentional harm, where a benign operator's agent misbehaves without anyone trying to hide it, plain lexical canaries — exact-match tokens — work well because the operator isn't actively hunting for and stripping them out. The harder case is intentional harm, where an adversarial operator might scrub anything that looks like a planted marker. For that setting, the authors propose semantic canaries tied to information the agent actually needs to complete its task, so removing the canary also removes what the agent needs to act. That creates what they call a utility-evasion asymmetry: an adversary who strips the signal can't stay useful.
They test the approach across communication, web, and cyber-agent settings and report that it produces reliable session attribution, holds up against adaptive attempts to transform away the signal, and scales to log searches vendors can realistically perform without being flooded with lookup requests.
What makes this notable against the current pile of agent-accountability proposals is the narrowness of the ask. It doesn't require a new identity standard, a new credential format, or vendor cooperation on provenance infrastructure. It asks vendors to do one thing they're already capable of: search existing session logs for a specific marker, in a specific time window, on request from an authorized party. The tradeoff, as outside reviewers have noted, is that the entire scheme rests on vendors actually granting that log-search access — a cooperation question the paper doesn't resolve, only assumes.