FULL STORY

OpenRouter's Leak Scandal: From Data Breaches to Provider Purge

OpenRouter suffered two data leak incidents within a month, casting doubt on its zero-retention promises. On November 22 it delisted three providers and removed 125 endpoints from its zero-data-retention pool.

2026-11-21 ~ 2026-11-23 · 2 episodes · 10 posts

Episode 1 · OpenRouter hit by two leaks in a month; MiniMax prompts possibly read by third parties (2026-11-21, 5 posts)

OpenRouter has suffered two data breaches within a single month, challenging its claimed data-retention policy of "monitoring every endpoint of every provider." On November 10, Phala, a hosting provider OpenRouter routes to, disclosed that an attacker had accessed data between October 21 and November (the exact end date was cut off in the source material). Then on November 21, developer Nate Berkopec made public a security notification from OpenRouter, which multiple users and researchers subsequently shared and confirmed.

Confirmed

  • OpenRouter's official notification stated that between July and October 2026, a security researcher planted unique test URLs (canaries) in prompts sent to MiniMax models; these URLs were accessed by a third party 40 to 130 hours after the requests were sent—regardless of the prompt content.
  • A notification email from OpenRouter Support shared by user bclavie showed that user requests routed through Together, GMI Cloud, and calls to MiniMax models "may have been read by the relevant party."
  • The attack disclosed by Phala on November 10 was the first leak of the month.

Not yet confirmed

  • The identity and purpose of the third party that accessed the data remain unclear, and the source material does not include further investigation conclusions from OpenRouter or MiniMax.

Why it matters

  • Security researcher Burkov pointed out that nearly all requests sent to OpenRouter should be treated as potentially readable by others—big players like OpenAI and Google at least have internal access controls, but when open-weight models are run on servers set up by individuals, prompts are very likely to be stored and analyzed.
  • bclavie also reiterated a security assumption: requests sent to third parties other than model labs or services like Bedrock should be assumed to carry a default risk of leakage.
  • The incident has already had real consequences: some developers, such as Nate Berkopec, have chosen to abandon and migrate away from OpenRouter.

Episode 2 · OpenRouter Offboards Three Providers, Removes 125 Endpoints from Zero-Data-Retention (2026-11-22, 5 posts)

On November 22, OpenRouter published a security bulletin: security researchers found evidence that some third-party inference providers on the platform had data leaks or had failed to honor their data-handling policies, though no evidence was found that OpenRouter itself was breached or accessed without authorization. It's a rare public handling of a supply-chain-level data security incident for an LLM aggregation routing platform.

Confirmed

  • OpenRouter officially off-boarded 3 providers and removed 125 endpoints from ZDR (zero data retention) eligibility, notifying affected customers.
  • The researchers' testing methodology will be incorporated into the platform's future security verification process.
  • Developer nateberkopec confirmed receiving an email from OpenRouter: providers Atlascloud, Novita, and Streamlake—labeled as US ZDR—were actually forwarding user data back to China's MiniMax, contrary to their commitments. Per m4, the off-boarded provider list is Atlas, GMI, and Phala, which conflicts with the former; the official follow-up disclosure should be treated as authoritative.

Why it matters

  • OpenRouter aggregates many model providers, and user trust in the ZDR label is a core platform asset; this incident shows that third-party data-policy commitments may not actually be enforced.
  • A view relayed by @irombie holds that such incidents show confidential computing is the fundamental path to compliant data handling—vendor self-discipline and policy statements alone aren't enough to earn user trust.
  • For enterprise users, choosing endpoints with a ZDR label doesn't exempt them from data-compliance scrutiny; they need to verify that the platform's validation mechanisms are actually in place.