SANS ISC
Dragos finds AI services leaking canary tokens in testing
SANS ISC’s Friday 20 November 2026 Stormcast covers several topics raised in the latest advisories and blog posts. The episode notes updates to PHP releases, highlighting memory-related issues that accompany remote code execution vulnerabilities in PHP’s core and in session management code. A non-critical quirk is also described: if a PHP file open operation connects to a remote URL using HTTPS but spells the protocol in uppercase, the code may fall back to HTTP.
The host says a detailed blog post with further analysis will be linked in the show notes, but stresses that this particular issue is not considered highly exploitable or critical.
In a separate thread, HD Moore (now with Dragos) used canary tokens—deceptive tokens that should appear stolen or misused when submitted—to probe AI tool ecosystems.
Over a million requests were analysed, and the findings suggested multiple leakage vectors: some AI providers exposed tokens due to internal compromises and connections to their AWS accounts; inference marketplace proxies inspected user requests and could see tokens; and a number of organisations appeared to accept requests while misrepresenting their privacy policies, effectively using data in prompts.
The host notes that canary tokens remain a useful, precautionary technique to detect data leakage, especially with AI services where data submission may be used for training or other purposes.
The episode closes with a mention that a vSphere vulnerability is reportedly being exploited, as described in a Google/Mandiant blog, with the host promising to share the URL in show notes for affected users. No CVEs or version numbers are provided in the transcript.