Dragos Finds Evidence of Intrusion at Three AI Providers and Data Leaks in Inference Marketplaces
·2026.11.19 22:00
Key point
A five-month canary token study revealed that three AI providers suffered intrusions and inference marketplace resellers accessed customer prompts despite Zero Data Retention settings.
Details
- Dragos researchers identified evidence of intrusions at three AI providers, where attackers accessed controlled test systems and canary AWS accounts, exposing data from at least three providers.
- Inference marketplace resellers accessed and stored customer prompts using standard logging features without exploiting vulnerabilities, with one case showing a home user attempting SSH login using a password from a prompt.
- Zero Data Retention (ZDR) policies were frequently ignored, with a US-based provider's prompts leaking to a Chinese AI organization, where secret URLs were scraped and shared in Feishu channels within 24 days.
- AI routers and marketplaces pose significant risks as they can view full prompts and responses at each hop; some sellers used stolen AWS credentials or consumer API keys to reduce inference costs.
Intrusion Details
- Attackers validated stolen keys within hours, then enumerated S3 buckets, accessed Secrets Manager, and enabled AWS Bedrock models, generating tens of thousands of API calls.
- One attacker compromised two providers simultaneously from a single rented server, using tools ranging from Windows Chrome scripts to Kali Linux AWS CLI and the LLM-based penetration testing framework ARTEX.
- Providers were unaware of the intrusions until Dragos notified them, with some initially misattributing the activity to internal configuration errors.
Supply Chain Risks
- Traceroute analysis of approximately 2,000 paths revealed unlisted intermediate layers, including a marketplace seller adding 1,000 hidden tokens to prompts and another using shared consumer subscriptions.
- Some providers failed to report the incidents publicly, and stolen keys and secret URLs remained accessible months after the initial compromise.
- Recommendations include self-hosting for sensitive tasks, using direct primary providers, avoiding reseller marketplaces, and implementing canary seeding and secret filtering in AI pipelines.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.