Skip to content
ARTEX.sh

Canary tokens for AI prompts:a tripwire for unexpected access to your context.

AI agents can send source code, tool output and credentials to model providers. A controlled canary gives defenders a signal when an item placed in that context is later accessed. This guide separates that signal from proof of who leaked it.

Checked

What a canary can show

Thinkst provides unique HTTP URLs that generate an alert when visited, and AWS key tokens that alert on API use. Put a token in one documented location so later access has a useful meaning. A canary is a tripwire; it does not prevent the original exposure.

Dragos used seeded prompts to investigate AI data handling. Its report also describes an attacker using ARTEX to check already stolen bait credentials. The incident page explains that observation without attributing the provider breaches to ARTEX.

Use a controlled test, not a real secret

The following is an operational checklist for a test you own. Use a purpose-built canary rather than a production credential. Keep customer data and working secrets out of the prompt.

  1. 01

    Define the question

    Name the agent, context source, model, router and upstream provider you intend to examine. Record the request identifier, privacy settings and time.

  2. 02

    Separate the tokens

    Assign a distinct token to each test route or context location. Record that mapping in a restricted ledger without publishing the token itself.

  3. 03

    Check the alert path

    Trigger a disposable test token yourself. Confirm who receives the alert and how the event is logged, then use a fresh token for the actual observation.

  4. 04

    Observe without creating access

    Keep the token inside your controlled material. Do not give the observer a working account, production permissions or a path into another organization.

An alert needs context

An HTTP alert establishes that the URL was requested. It does not by itself distinguish an attacker from a link preview, safety scanner, retrieval tool or your own test. Record expected automated fetches before drawing conclusions.

Thinkst notes that AWS token notifications can be delayed by Amazon logging. Compare event times and request identifiers rather than treating notification time as the moment of access.

A token that passes through multiple services cannot identify which hop exposed it. Keep the router and provider trail; compare distinct controlled routes before making an attribution. No alert is not proof that a prompt was never stored or copied.

Preserve the evidence and reduce exposure

Save the alert, relevant request identifiers, timestamps and redacted logs. Restrict access to the ledger. If a real secret may also have been exposed, follow your incident process to revoke it and inspect its use. Contact the relevant provider privately with the evidence.

Canaries complement secret filtering, limited agent access and deliberate provider selection. They do not certify a provider, prove compliance with retention settings or make it safe to send real secrets. This workflow is editorial guidance, not a test result from this site.

Questions

Does a canary alert prove an AI provider was breached?
No. It proves the token was accessed in the way that token detects. Expected fetchers, the request route and corroborating logs determine what you can conclude.
Can I use my real AWS key as bait?
Use a purpose-built token instead. A production key creates real access and turns an observation into an avoidable exposure.

Sources

Every claim on this page comes from these sources or from this site's own sourced pages. Open them for the full reports.

Follow the ARTEX story.

New forks, releases and reports, dated and sourced.

Isometric plane with glowing light columns