Dragos exposes prompt leaks, provider breaches
Dragos seeded two million AI prompts with trackable canary tokens, discovering active threat actor compromises at three hosting providers and rampant prompt logging by marketplace resellers. The investigation revealed that Zero Data Retention policies are routinely bypassed, with confidential prompts and credentials leaking to downstream intermediaries and external scrapers.
The illusion of end-to-end privacy in AI routing is dead: if you route sensitive prompts through cheap aggregators or third-party inference platforms, your data and API credentials will eventually leak. Zero Data Retention settings offer zero defense once an attacker sits inside a provider's ingress layer or an intermediary logs requests to undercut GPU spot prices.
- –Attackers stole canary AWS keys before requests reached target models, triggering thousands of unauthorized API calls and confirming that provider front-door proxies were breached.
- –Discount inference resellers routinely harvest prompts from platform logs, with individuals on residential connections observed attempting manual SSH logins using credentials extracted from prompts.
- –Zero Data Retention flags failed across multiple hops, with test links forwarded internationally, crawled weeks later by persistent scrapers, and leaked into internal company chat channels.
- –Developers using AI agents or routing proxies must assume prompt context is public; teams should scrub credentials from prompts, deploy canary tokens to audit suppliers, or self-host models for sensitive data.
DISCOVERED
2h ago
2026-11-22
PUBLISHED
5h ago
2026-11-22
RELEVANCE
AUTHOR
kristianp