AI agents can be unleashed (intentionally or unintentionally) onto a network looking for vulnerabilities and sensitive data with speed and tenacity beyond that of human attackers. Canaries are well-positioned to detect these agents poking around in your network.
Your Canary can be configured with one of the following personalities to allow it to detect interactions from AI agents: Agent Provocateur or Agent-to-Agent (A2A).
These personalities pose as services that can provide assistance to agents interacting with them, but instead of actually helping the agent with its objectives they instead ask the agent questions about its task, identity and the environment its running in.
The Agent Provocateur service is self-describing web page that points an agent to its API where it can interact with subagents, such as an environmental support agent, a result validator and more. On the other hand, the Agent-to-Agent service interacts with the agent through the A2A Protocol instead of a bespoke API, while still querying the agent for information.
Note
These personalities don't try to mimic existing services and instead try to extract information out of agents interacting with them. As such, it's possible that they could be fingerprinted and these services should therefore be run on standalone Canaries (not simply enabled on your existing Canaries which are configured to mimic real services).
Examples of the alerts generated by the Agent Provocateur and Agent-to-Agent services are shown below.