Estimated reading time: 7 minutes · Last updated:
AI prompt theft ran for months inside three AI inference providers, and none of them knew until researchers at Dragos, an industrial cybersecurity company, told them. Between July and November, Dragos planted unique links and credentials in about two million prompts, then watched who used them. Some stolen requests carried Zero Data Retention, a setting meant to guarantee a prompt is never stored. On 22 November 2026, OpenRouter, a gateway that passes customers' AI requests on to a roster of third-party model hosts, said it had removed three providers, pulled 125 endpoints from Zero Data Retention eligibility and notified potentially affected customers. OpenRouter said its own systems were not breached.
A privacy flag doesn't help when the attacker is already inside.
HD Moore, chief research officer, Dragos
Key takeaways
- Months of undetected access: Dragos traced attackers inside three AI inference providers, the earliest from mid-July, using bait seeded in about two million prompts between July and November.
- OpenRouter's cut: OpenRouter dropped three providers and stripped 125 endpoints of Zero Data Retention eligibility, and says any endpoint wanting that status back must first pass assurance and validation.
- Privacy flags ignored: An unnamed Chinese AI lab fetched secret links from roughly two-thirds of seeded prompts marked Zero Data Retention or no data collection that reached its models.
- Phala exposure window: Phala says DeepSeek V4 Flash 0731 requests sent via OpenRouter between 21 October and 8 November 2026 count as exposed, and secrets in them should be rotated.
- Together AI leak: Together AI says a credential pushed to its public GitHub repository by an AI coding agent let an outsider download prompt samples from June and July.
Table of contents
- Key takeaways
- How Two Million Bait Prompts Caught Thieves Inside Three Hosts
- How Sellers and a Chinese Lab Collected Prompts Without a Break-In
- What Phala, Together AI and Novita Have Disclosed
- What OpenRouter Changed and What It Has Not Said
- Will AI Prompt Routing Get Safer
- What to be careful about
- Frequently asked questions
How Two Million Bait Prompts Caught Thieves Inside Three Hosts
A canary token is a fake secret, such as a link or password, that exists only inside one prompt. Anyone who uses it must have read that prompt. The earliest breach Dragos uncovered began in mid-July, and because the bait was processed alongside genuine customer requests, Dragos said those were almost certainly copied too.
By Dragos's account, the largest breach centred on a couple hundred planted AWS keys, each sent once to a single provider during a few brief windows in late October. Every request in those windows was taken, including some the provider rejected before any model ran. Dragos read that as a sign the attacker copied traffic where it first arrived.
Some keys were tried within hours. The intruder enumerated S3 buckets, combed Secrets Manager and Parameter Store, and tried to mint Bedrock API keys and add IAM users; the activity amounted to tens of thousands of calls against AWS's API. Dragos believes a single attacker, working from one rented server, held two providers simultaneously. On November 1 that attacker ran ARTEX, an AI-driven penetration testing framework, purely to check that stolen keys still worked.
How Sellers and a Chinese Lab Collected Prompts Without a Break-In
Dragos's second finding needed no intrusion. Sellers on an AI inference marketplace undercut rivals on price, turned on request logging and kept what passed through. They looked like individuals on home broadband. One, on a Mac in France, lifted passwords from two prompts, tried root SSH logins by hand and tested several planted AWS keys.
One key had reached the marketplace through a separate router that forwarded some traffic there without telling customers; that router has told Dragos it stopped. Tracing about 2,000 routes found several hundred unidentified systems and a seller that slips roughly 1,000 hidden tokens into each prompt.
A Chinese AI lab that Dragos does not name retrieved the secret links in over half the seeded prompts its models received between July and October, including about two-thirds flagged Zero Data Retention or no data collection. U.S. hosts reselling its models set off the same scraper. Later, a git client on a separate cloud attempted to clone private repositories named in prompts, for that lab's models alone, and a large U.S. provider was still crawling one secret link over two months after a single query to its search endpoint.
Twenty-four days after one prompt went out, its link was posted to a Feishu chat; Feishu is the Chinese workplace app that goes by Lark elsewhere. Three desktop computers and three phones across Beijing and Shanghai opened it within two minutes. Dragos had made public just the first 8 characters, so the poster held a cached copy of the full prompt. HD Moore, Dragos's chief research officer, wrote that flagged prompts were kept, shared with other companies and read by people.
What Phala, Together AI and Novita Have Disclosed
Phala said last week, in a preliminary incident report, that an attacker controlled one of its inference servers and could see request data mid-processing. It contained the intrusion on November 8 and restored service with added safeguards. Phala says DeepSeek V4 Flash 0731 requests reaching it via OpenRouter from 00:00 UTC on October 21 to 05:35 UTC on November 8 should be treated as exposed, with any secrets in them rotated. A researcher had seeded canaries since July 1, and only those planted on October 21 and 22 were misused, but that sampling missed some requests, so Phala kept the broader window.
Together AI said on 21 November 2026 that an outsider downloaded stored prompt samples after an AI coding agent pushed a credential into its public GitHub repository. The samples, kept for routine software testing, came from June and July requests to MiniMax M3, GLM-5.2 and Kimi K2.6, and five credentials the intruder tested were Dragos canaries. Together AI believes the data went no further, understands it was deleted, and now starts new accounts on zero data retention. Novita confirmed a routing configuration issue sent some MiniMax requests on to MiniMax's official API, beyond what customers intended.
What OpenRouter Changed and What It Has Not Said
OpenRouter said the 125 endpoints it pulled from Zero Data Retention eligibility include providers Dragos flagged along with others dropped as a precaution, and none will regain that status without assurances and validation. It is turning Dragos's test into an ongoing check and verifying network paths to see where requests and responses might be forwarded to third parties. Not every affected provider has spoken publicly; OpenRouter has notified users regardless, and three providers have posted.
Neither Dragos nor OpenRouter has identified its three, so the public cannot tell whether the lists overlap. The provider breached in mid-July first pointed to a misconfiguration inside the company, then to how a partner handled data, before its investigation showed attackers had taken its internal credentials. As of this week it had disclosed nothing.
Roughly a dozen providers are now working with Dragos, and as of this week someone was still touching the decoy keys and secret links. In October, the platform company Railway said it had likely been hit as part of a string of attacks on platform companies, with AI inference providers as the eventual target.
| Provider | What happened | Models named | Window | Stated response |
|---|---|---|---|---|
| Phala | Attacker controlled an inference server | DeepSeek V4 Flash 0731 | 21 Oct to 8 Nov 2026 (UTC) | Contained 8 Nov; rotate secrets |
| Together AI | Prompt samples downloaded after leaked GitHub credential | MiniMax M3, GLM-5.2, Kimi K2.6 | June and July 2026 | Zero data retention default for new accounts |
| Novita | Routing issue sent requests to MiniMax's official API | MiniMax models | Not stated | Privacy notice issued |
Will AI Prompt Routing Get Safer
The case for
- OpenRouter is converting Dragos's canary testing into a continuous check and verifying network paths for undisclosed forwarding.
- Endpoints removed from Zero Data Retention must pass assurance and validation before regaining it.
- Together AI has made zero data retention the default for new accounts.
- The router that forwarded traffic to the marketplace has told Dragos it stopped.
The case against
- The provider breached in mid-July had made no public disclosure as of this week.
- Neither Dragos nor OpenRouter has named the providers involved, so customers cannot match exposures to vendors.
- Planted keys and secret links were still being accessed this week, months after the first thefts.
- Route tracing found several hundred unidentified systems and a seller adding about 1,000 hidden tokens to prompts.
What to be careful about
- Credentials embedded in prompts sent through the affected providers may still be in attackers' hands; Dragos saw planted keys used this week.
- A Zero Data Retention flag governs only the first system a request reaches, so downstream resellers and labs can still retain prompts.
- Reseller marketplaces may run partly on stolen keys and can log requests without customers knowing.
- Filters that strip credential formats from prompts miss sensitive data that does not look like a key.
The bottom line
For customers, Moore's guidance begins with a limit: a privacy setting governs only the first system a request reaches, and AI agents forward anything they can read, secrets included. In Dragos's view, self-hosting is the only setup where no outside party sees a prompt, and going straight to a first-tier provider leaves a single company, contract and privacy policy to vet. Router users should pick upstream providers themselves, test whether privacy settings really change a request's route, and avoid reseller marketplaces, some of whose capacity relies on stolen keys. A decoy credential planted in the context agents already send, with any use treated as an incident, turns the method that exposed these breaches into a standing alarm.
What to watch
- Watch for OpenRouter to restore Zero Data Retention status to any of the 125 removed endpoints once they pass its assurance and validation checks; no date has been set.
- Watch for the provider breached in mid-July to disclose the incident publicly; no date has been set.
- Watch for Dragos or OpenRouter to name the providers involved or confirm whether their two lists of three match; no date has been set.
Frequently asked questions
Does Zero Data Retention stop AI providers from keeping my prompts?
Not by itself, on Dragos's evidence. The setting is meant to ensure a prompt is never stored, yet an unnamed Chinese AI lab fetched secret links from about two-thirds of seeded prompts carrying that flag or a no data collection flag. HD Moore of Dragos wrote that such a flag offers no protection once an attacker is inside.
Which ARTEX user agents should security teams add to detections?
Dragos named two: artex-agent and artex-judge-r121. They come from ARTEX, an AI-driven penetration testing framework one attacker used on November 1, and Dragos says neither appears on any public indicator list it could find.
How can a company set a canary trap for AI prompts?
Dragos suggests planting a decoy credential in the data that agents and pipelines already transmit, and treating any use of it as an incident. A free Thinkst Canarytoken is a workable starting point. Stripping recognisable credential patterns from prompts adds a layer, though it misses sensitive data that doesn't look like a key.
Related reading