Published
- 5 min read
By Allan D - Editor, AI Security Wire
OpenRouter Drops 3 Providers After Dragos Finds Stolen AI Prompts
A security research team seeded two million AI prompts with fake secrets and watched who used them. The answer, per Dragos, is that attackers broke into AI hosting providers, resellers read customer prompts through ordinary logging, and privacy settings did not stop prompts travelling further than customers expected. On 22 November, OpenRouter said it had off-boarded three providers and removed 125 endpoints from Zero Data Retention (ZDR) eligibility.
Everything below is the companies’ own account. As of 22 November, in the five sources we opened, we found no independent forensic report. Together and Phala confirm their own incidents, and Together ties its incident to Dragos canaries. Novita says it observed no use of the AWS test key markers. Neither Dragos nor OpenRouter names the three breached providers.
What Dragos did
Dragos published the research on 19 November, written by Chief Research Officer HD Moore. From July to November 2026 it sent roughly two million prompts to widely used AI services. Each carried unique URLs and canary credentials, including AWS keys tied to empty, fully logged accounts. Any later use of those secrets showed that someone had seen the prompt.
Dragos reports three findings. First, data from at least three providers was exposed. The earliest incident dates from mid-July, and none of the providers knew until Dragos notified them. The largest involved a couple of hundred AWS keys sent in a few windows in late October, and every request in those windows was stolen. Some stolen requests were ones the provider rejected before a model ran, which Dragos reads as copying at the provider’s front door. In the other two cases, some stolen requests were flagged ZDR.
Once attackers had the keys, Dragos says attackers tested them, in some cases within a few hours of Dragos’s request, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to enable Bedrock models and add their own IAM users. The bait keys saw tens of thousands of AWS API calls. One adversary appears to have compromised two providers at once. On 1 November that attacker added ARTEX, an LLM-driven pentest framework, which Dragos says made “a handful of calls to confirm that keys the attacker already held were still live”. Its user agents, artex-agent and artex-judge-r121, are worth adding to detections. We covered the ARTEX framework family earlier this month.
Second, resellers on an inference marketplace read and kept prompts using standard logging, with no exploit needed. Third, Dragos says providers “appeared not to honor customer privacy settings or published policies”. Systems at one Chinese AI lab fetched the secret URLs from more than half the prompts sent to its models between July and October, including roughly two-thirds of those marked “no data collection” or ZDR, and U.S. providers reselling its models triggered the same scraper.
What providers have said
OpenRouter’s post says the findings “did not show any breach or unauthorized access of OpenRouter.” It says it off-boarded three providers, removed 125 endpoints from ZDR eligibility (some as a precaution), notified potentially affected customers, and will require validation before restoring ZDR. It links three provider notices, and says not all affected providers have posted publicly.
- Novita (names no researcher): a routing configuration issue sent some MiniMax requests to MiniMax’s official API instead of Novita infrastructure. Requests between 28 May and 11 November 2026 may have been accessed or retained by a third party, including those routed through OpenRouter. Novita suspended the models and says it observed no use of the AWS test key markers.
- Phala (preliminary report, 10 November): an attacker chained a legacy internet-facing service with management interfaces that trusted loopback access, then took over an inference VM. Phala says DeepSeek V4 Flash 0731 requests from 21 October to 8 November should be treated as exposed. Phala says only canary credentials from requests sent on 21 and 22 October were subsequently misused, and it does not name Dragos. It was contained on 8 November, and a full root cause report is promised.
- Together AI (affected requests from June and July 2026; review ongoing): stored prompt samples from three models were downloaded after an AI coding agent committed a credential for a private Hugging Face repository to a public GitHub repo. Together says five of the credentials tested were canary credentials associated with a Dragos security experiment, and it is making ZDR the default for new accounts.
No source says which providers are the three Dragos describes as breached or the three OpenRouter off-boarded. OpenRouter lists these three notices but does not say they were the ones flagged, and only Together’s notice names Dragos. OpenRouter ties its off-boardings to “the results of our investigations”, not to the Dragos report. Treat any mapping, including to the Chinese AI lab in Dragos’s third finding, as unconfirmed.
Why it matters
This is a supply chain problem for anyone sending sensitive material through AI routers. Agents will paste whatever they can read, credentials included. Dragos says a privacy flag only binds the first hop, and advises picking upstream providers yourself instead of letting the lowest price decide.
For your own exposure, search logs for traffic through the named providers and windows, and rotate anything secret that was in those prompts. Dragos also recommends seeding fake credentials as canaries in agent and pipeline context so leaks surface early.
References
- Dragos — Three breaches, endless leaks: Your AI supply chain needs canaries
- OpenRouter — Data Processing Incidents on Inference Providers
- Novita AI — Privacy Notice Regarding MiniMax Requests on Novita
- Phala — November preliminary incident report
- Together AI — Security event: prompt leakage for requests in June and July of 2026
Frequently Asked Questions
- What did Dragos find?
- Dragos sent about two million prompts containing unique canary URLs and credentials to AI services between July and November 2026 and watched where they surfaced. It reports evidence that data from at least three AI hosting providers was stolen, that marketplace resellers read and kept customer prompts, and that providers 'appeared not to honor customer privacy settings or published policies'. None of the three breached providers knew until Dragos told them, according to Dragos.
- Does Zero Data Retention protect my prompts?
- Not reliably. Dragos says some stolen requests were marked as Zero Data Retention, and that a privacy flag only binds the first hop in a chain of routers and providers. OpenRouter has removed 125 endpoints from its ZDR eligibility pending validation.
- What should teams using routed AI inference do?
- If your requests went through the affected windows, rotate any credentials, keys or tokens that appeared in prompts. Going forward, choose upstream providers yourself instead of accepting a default pool, as Dragos advises, and consider Dragos's suggestion of seeding canary credentials in agent and pipeline context to detect leaks. Our advice: keep secrets out of prompts where you can.