enzosoler.comreviews 09bench 9 deviceslive test Can a ring and a watch replace my Apple Watch?
Published

OpenRouter dropped three AI providers after a study caught prompts leaking. Rotate any secret you sent through them

Close-up of fiber optic cables plugged into a network switch in a data center
Image: Brett Sayles / Pexels

A five-month Dragos study planted fake credentials in about two million AI prompts and found attackers stole data from at least three inference providers, some of it from requests marked Zero Data Retention. OpenRouter removed three providers, pulled 125 endpoints from ZDR eligibility and notified affected customers. Phala and Together AI posted their own incident reports.

If your company sends prompts to open-weight models through OpenRouter or another AI router, check your inbox for a notice from OpenRouter, then rotate any API key, password or token that ever appeared in those prompts. On November 22 OpenRouter said it off-boarded three inference providers and removed 125 endpoints from Zero Data Retention (ZDR) eligibility after security researchers showed that some providers had been breached or were not keeping their data promises.

What the research found

Dragos chief research officer HD Moore published the study on November 19. From July through November, his team sent about two million prompts that each carried a unique link and fake credentials, then watched who used them.

  • Three providers breached. Attackers stole request data from at least three hosting providers, starting in mid-July. Dragos says none of them knew until it told them.
  • Stolen keys used within hours. In the largest breach, a couple hundred planted AWS keys were stolen and used for tens of thousands of AWS API calls.
  • ZDR did not stop it. In two of the three breaches, some stolen requests were marked ZDR. A Chinese AI lab’s scraper fetched the secret links from about two thirds of requests flagged “no data collection” or ZDR.
  • Hidden hops. Dragos traced about 2,000 request routes. Several hundred passed through a system nobody had disclosed.

Dragos does not name the three breached providers, and OpenRouter does not name the three it dropped. OpenRouter says it found no breach of its own systems. According to The Hacker News, neither company has said whether the two groups of three are the same.

Who has confirmed an incident

  • Phala. Its November 10 report says an attacker compromised a GPU server running DeepSeek V4 Flash 0731, reached only through OpenRouter. Treat any request to that model sent between October 21 and November 8 at 05:35 UTC as exposed.
  • Together AI. On November 21 it reported that an AI coding agent pushed a credential to a public GitHub repo, and an attacker used it to download stored prompt samples from June and July. Affected requests went to MiniMax M3, GLM-5.2 and Kimi K2.6. ZDR is now the default for new Together accounts; existing accounts must switch it on themselves.
  • Novita. It says a routing setting sent some MiniMax requests on to MiniMax’s own API.

What to do now

  • Rotate at the source. Revoke every secret that went into a prompt through a router or a reseller, starting with requests to the models and dates above.
  • Pick your upstream providers. In your router settings, choose which providers may serve your requests. Do not accept the default pool, and avoid reseller marketplaces for anything sensitive.
  • Turn on ZDR where you can, and test it. Dragos found that a privacy flag often covers only the first hop. Check that changing it changes where your requests go.
  • Plant a tripwire. Put a fake credential, such as a Thinkst Canarytoken, in your agent and pipeline context. Treat any use of it as an incident.
  • Set spend caps at every AI provider, so a stolen key hits a ceiling.

The cheapest model route now carries a vendor risk

Routers made it easy to send a prompt to whichever host was cheapest or fastest that minute. This study shows the cost of that choice: each extra host is another company holding your data, and your ZDR setting only binds the first one. For prompts that carry customer data or credentials, Dragos recommends one first-tier provider with one contract and one privacy policy, or self-hosting.

The same pattern ran through October’s breaches at AI developer platforms. Railway says stolen AI keys ran up about $740,000, and Notion traced exposed AI chat transcripts to a breach at its model vendor. Dragos says one attacker used ARTEX, the AI hacking framework the FBI warned about on November 13, and suggests adding its user agents “artex-agent” and “artex-judge-r121” to your detections.

Related, November 22: Cleverbridge says an attacker copied card numbers and CVV codes from its software checkout platform. If you bought a license through it, check your statements and consider a new card.

One email when a review goes up. Unsubscribe from any of them.