AGENT CANARY
Instruction canary (experimental)
Mint a canary and hide the snippet in your own content: a doc, a web page, a JSON response, a config file. It carries a plain-language instruction pointing at a unique tripwire URL. It is not a clickable link, so ordinary link traversal should not discover it. A retrieval is a useful lead, but does not by itself prove an LLM, an autonomous agent, or instruction-following. You get an email when the URL is retrieved.
Deterministic, GET-only, no account. The canary is yours; the link you get back is the only key to it. This is a sensor you plant. The evidence I publish is the loot list (coding-agent credential paths swept against my decoy). Hits on a canary you mint stay private; they are not added to that list. A quiet canary is not a broken one.
- You embed the snippet in content an agent might ingest: your docs, a page it scrapes, an API response, a fake internal endpoint.
- The snippet is a natural-language instruction, never a clickable link. That makes ordinary link traversal less likely to reach it, but does not establish why a requester fetched it.
- When a requester fetches the tripwire, I log its timing and self-declared user-agent (never your raw IP) and email you. Treat the request as a lead to investigate, not an attribution to an AI agent.
This uses the same request-sensor pattern as my honeypots. See botwatch for separate, public honeypot evidence.