OpenRouter removed three AI providers on November 22 after Dragos reported months of prompt theft at three inference firms. The gateway also stripped 125 endpoints of eligibility for Zero Data Retention and notified all potentially affected customers.
Dragos said attackers took requests carrying that privacy setting, which is supposed to prevent prompt storage. Its researchers also found undisclosed forwarding, retained prompts and people reading data customers had submitted to AI services.
OpenRouter routes AI requests to third-party providers. It said the findings showed no breach of its own systems.
The three breached firms were unaware of the intrusions until researchers alerted them, according to Dragos. The earliest compromise began in mid-July. Researchers said real customer requests processed alongside their test prompts were almost certainly copied too.
Neither Dragos nor OpenRouter named its three providers. They have not publicly established whether the breached firms and removed providers are the same companies.
Provider notices identify requests at risk
Three providers have published disclosures, although others affected have not, OpenRouter said. The gateway said it notified affected users regardless of whether their provider issued a public notice.
Phala’s preliminary report described an attacker gaining control of an inference server. That access allowed the attacker to see requests during processing.
Phala advised treating requests to DeepSeek V4 Flash 0731 through OpenRouter as exposed during a specified period. The window runs from October 21 at 00:00 UTC to November 8 at 05:35 UTC.
Customers should rotate every key, token, password or other secret included in those requests, Phala said. It contained the incident on November 8 and restored the model service with additional protections.
A researcher had submitted unique test credentials since July 1. Only credentials sent on October 21 and 22 were later misused. Phala retained the wider exposure window because those tests sampled traffic rather than covering every request.
Together AI disclosed a different route to prompt theft in its November 21 security notice. An AI coding agent committed a credential to the company’s public GitHub repository.
An unauthorized party then downloaded stored customer prompt samples gathered for routine software testing. Those samples came from June and July requests to MiniMax M3, GLM-5.2, and Kimi K2.6.
The party searched for credentials to other services and tested them. Five tested credentials were canaries planted by the researchers.
Together AI said it had no reason to believe the data spread further and understood that it had been deleted. It has made zero data retention the default for new accounts.
Novita’s privacy notice described a routing configuration issue rather than an intrusion. Some MiniMax requests went to MiniMax’s official API beyond what those customers intended.
Planted secrets revealed access to prompts
Dragos’s research used about two million prompts containing unique links and credentials between July and November. Researchers watched for visits to those links and attempts to use the credentials.
Each secret appeared in a specific prompt. Its later use therefore revealed that someone had accessed that prompt.
The largest breach exposed a couple hundred planted AWS keys during several brief periods in late October. Each key had appeared in exactly one request to one provider.
Dragos said every request submitted during those periods was stolen. Some had been rejected before a model ran. That suggested copying where traffic first entered the provider, rather than later during inference.
Some stolen keys were tested within hours. The attacker listed S3 buckets and searched Secrets Manager and Parameter Store.
Attempts to create Bedrock API keys and add IAM users sought to preserve access. The activity generated tens of thousands of AWS API calls.
One attacker appeared to access two providers, using secrets from both through a single rented server. On November 1, that attacker used ARTEX, an AI-driven penetration testing framework, to check whether previously obtained keys still worked.
Dragos identified the user agents artex-agent and artex-judge-r121. It found neither in public indicator lists and recommended adding them to detection rules.
“A privacy flag doesn't help when the attacker is already inside,” wrote HD Moore, Dragos’s chief research officer.
Logging and hidden routes exposed more data
Other findings required no intrusion. Dragos found inference marketplace sellers attracting requests with the lowest prices, enabling logging and collecting the resulting traffic.
The operators appeared to be individuals using home internet connections. One used a Mac on home broadband in France to manually attempt root SSH logins with passwords from two prompts.
The same machine tested several planted AWS keys. One key had passed through another router that forwarded some requests to the marketplace without telling customers.
That router later told Dragos it had stopped the forwarding. Tracing about 2,000 routes uncovered several hundred unidentified systems.
Researchers also found one seller adding about 1,000 hidden tokens to every prompt.
An unnamed Chinese AI lab fetched secret links from more than half the seeded prompts submitted to its models between July and October. That included about two-thirds of requests marked no data collection or Zero Data Retention.
Requests through U.S. providers reselling those models triggered the same scraper, showing that prompts flowed back to the lab.
A git client on another cloud later attempted to clone private repositories named in prompts. Those attempts occurred only for that lab’s models.
Separately, a large U.S. provider continued crawling a secret link more than two months after one request to its search endpoint.
Another link appeared in a Feishu chat 24 days after researchers submitted the prompt containing it. Feishu is known outside China as Lark.
Within two minutes, three desktops and three phones in Beijing and Shanghai opened the link. Dragos had disclosed only its first 8 characters, indicating that the person posting it had a cached copy of the full prompt.
Moore said the findings showed flagged prompts being retained, shared with other companies and read by people.
OpenRouter expands checks as investigations continue
OpenRouter’s incident announcement said the 125 endpoints included researchers’ findings and additional precautionary exclusions. Restoring Zero Data Retention eligibility will require assurances and validation.
The gateway is adapting the researchers’ tests for continuous use. It is also checking network paths to identify third parties receiving requests or responses.
Dragos is working with about a dozen providers. As of this week, planted keys and links were still being accessed months after the earliest thefts.
The provider breached in mid-July initially attributed the problem to an internal misconfiguration, then to a partner’s data processing. Its investigation later found compromised internal credentials, Dragos said. It had not publicly disclosed the incident as of this week.
Railway said in October that its own intrusion likely belonged to a series targeting platform companies. The suspected ultimate targets were AI inference providers.
Researchers advise choosing routes and testing privacy claims
Moore cautioned that a privacy flag governs only the first hop. AI agents can transmit credentials and anything else they can read.
The researchers described self-hosting as the only option that avoids a third party reading prompts. Using a first-tier provider directly leaves one company, contract and privacy policy to assess.
For router users, they recommended selecting upstream providers rather than accepting the default pool. Customers should also check whether privacy settings actually change routing.
They advised avoiding reseller marketplaces, saying some capacity operates on stolen keys.
A fake credential placed in the context agents and pipelines already transmit can reveal later access. Researchers said a free Thinkst Canarytoken could serve this purpose, with any use treated as an incident.
Prompt filters that remove known credential formats can also help. They do not catch sensitive information that lacks a recognizable key format.

