The Hacker News
Researchers uncover months of prompt theft at AI providers
Three AI inference providers were found to have had months of prompt theft, according to Dragos, with OpenRouter subsequently removing three providers from its platform and revoking 125 endpoints from its Zero Data Retention eligibility. Dragos’ researchers say the intrusions occurred over a period from mid-July, with planted prompts and credentials able to be copied along with real customer requests. OpenRouter stressed there was no breach of its own gateway, and that affected customers have been notified.
The investigation relied on canary-token research seeded across roughly two million prompts between July and November. A notable finding was that some stolen prompts contained credentials that allowed access to third-party services, including hundreds of AWS keys used in late October; attackers reportedly tested keys within hours, enumerated S3 buckets, interrogated Secrets Manager and Parameter Store, and attempted to create Bedrock API keys and IAM users, generating tens of thousands of AWS API calls.
One actor appeared to compromise two providers from a single rented server, and an AI-driven pentesting framework named ARTEX was used to verify live keys. The researchers note that in some cases, prompts flagged as “no data retention” were processed in a way that allowed adversaries to read the content.
OpenRouter said it has updated its screening to run ongoing checks, is tracing network paths to identify forwarded requests, and is working with about a dozen providers; three had disclosed incidents publicly, while Dragos has not named the breached providers. Some affected clients have posted notices; OpenRouter has warned users that any removed endpoints require validation before restoration of status.