AI inference providers: attackers stole prompts and AWS keys, including Zero Data Retention requests
Severity: high · Patch status: partial · Since 2026-11-10
Dragos sent about two million prompts with unique URLs and canary credentials between July and November 2026 and reports that attackers stole data from inside three hosting providers for months, none of which knew until Dragos notified them. The largest breach took a couple hundred planted AWS keys, each sent in one request during late-October windows; the attackers tested the keys within hours, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock API keys and IAM users, driving tens of thousands of AWS API calls. In two of the breaches some stolen requests were marked Zero Data Retention. Sellers on an AI inference marketplace also logged request traffic, and an unnamed Chinese AI lab fetched the secret links from more than half of the seeded prompts sent to its models. Phala reports that an attacker took control of an inference VM serving DeepSeek V4 Flash 0731 and advises treating requests sent 2026-10-21 00:00 UTC to 2026-11-08 05:35 UTC as exposed, a precautionary window its canary tests did not fully confirm. Together AI reports that an unauthorized party downloaded stored June and July 2026 prompt samples after a repository credential was committed to a public GitHub repository. OpenRouter off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility, and said its findings showed no breach or unauthorized access of OpenRouter itself. Neither Dragos nor OpenRouter has named the three breached or off-boarded providers.
What to do
- Rotate every API key, token, password, and secret that appeared in requests sent to Phala's DeepSeek V4 Flash 0731 between 2026-10-21 00:00 UTC and 2026-11-08 05:35 UTC, and in requests sent to any OpenRouter endpoint removed from Zero Data Retention eligibility.
- Audit AWS CloudTrail and IAM for S3 bucket listing, Secrets Manager and Parameter Store queries, Bedrock API-key creation, and new IAM users; revoke any unauthorized access, and treat every key that appeared in a prompt as compromised.
- Treat Zero Data Retention as not guaranteed for affected endpoints; review provider logging, routing, and any third-party forwarding.
- Remove or revalidate endpoints that lost ZDR eligibility before resuming sensitive prompts; OpenRouter says restoration needs assurances and validation first.
- Pin upstream providers yourself instead of accepting the default pool, verify that privacy settings change where requests actually go, and avoid reseller marketplaces (Dragos's routing advice).
- Notify affected customers and monitor canary links and credentials for later misuse.
Sources
All live incidents on Threat Watch