Definition
An attack that exploits a flaw in the rules an AI agent uses to decide who it should trust or obey. For example, an AI agent might be configured to only accept instructions from users in a whitelist — but if that whitelist checks a field that an attacker can change (like a display name), the attacker can spoof a trusted identity and issue unauthorised instructions.
Why it matters
Many AI agent deployments rely on simple, metadata-based checks to enforce trust boundaries. Research found this pattern broken across multiple messaging platforms simultaneously, meaning attackers in those channels could redirect agents' actions without needing any technical exploit.