Attackers stole prompts and credentials from inside three AI inference providers for months, and none of the three knew until security researchers told them, the researchers said this week. They seeded about 2 million prompts with unique web links and credentials between July and November and watched for use. The earliest of the three breaches began in mid-July, and some stolen requests were flagged for zero data retention. OpenRouter, the routing platform that sends customer requests to such providers, removed the three from its platform and cut 125 endpoints from zero-data-retention eligibility until they are validated. It notified customers who may be affected and said its own systems were not breached.
The largest breach
A couple of hundred AWS keys were planted in short windows in late October, and every request in those windows was stolen, including some the provider had rejected before any model ran. The thief tested keys within hours, listed storage buckets, searched secrets stores and tried to create new Bedrock keys and IAM users, which produced tens of thousands of AWS calls. One attacker appears to have compromised two providers from one rented server, and on 1 November used an AI-driven penetration-testing framework called ARTEX to confirm held keys were still live. Researcher HD Moore said a privacy flag does not help when the attacker is already inside.
Marketplaces and an AI lab
Sellers on an AI inference marketplace listed models at the lowest prices, switched on request logging and collected what flowed through. A Chinese AI lab fetched the secret links from more than half the seeded prompts sent to its models between July and October, including about two-thirds of those marked as not to be collected, and the same scraper fired when prompts went to US hosts reselling its models. In one case a link posted in a Feishu chat 24 days after the prompt was sent was opened within two minutes from desktops and phones in Beijing and Shanghai.
Providers' disclosures
Phala said an attacker controlled an inference server and saw request data, contained it on 8 November, and advised treating requests to DeepSeek V4 Flash that reached it between 21 October and 8 November as exposed. Together AI said an AI coding agent committed a credential to a public GitHub repository and an unauthorised party downloaded stored prompt samples from June and July requests, and that zero data retention is now the default for new accounts. Novita said a routing fault sent some MiniMax requests on to MiniMax's own API. Researchers are working with about a dozen providers. They advise self-hosting where no third party may read a prompt, going directly to a first-tier provider, avoiding reseller marketplaces and using fake credentials as tripwires.









