Three AI Providers Breached, None Knew Until Told

Attackers quietly siphoned private AI prompts from inside three inference providers for months, and not one of the three detected the intrusion on its own. The breaches only came to light after the industrial cybersecurity firm Dragos planted millions of canary tokens across AI platforms and watched for anyone who touched them. The findings have triggered sweeping action from OpenRouter, one of the largest AI routing gateways, which today announced it has removed three providers from its platform and stripped 125 endpoints of their Zero Data Retention (ZDR) status.

What Happened

Between July and November 2026, a team led by Dragos Chief Research Officer HD Moore sent approximately two million prompts to widely used AI services. Each prompt was seeded with a unique URL and a set of canary credentials—fake secrets that had never been used anywhere else. The logic was elegant: if anyone ever accessed one of those URLs or tried to use those credentials, it could only mean the prompt had been read by an unauthorized party after the model finished processing it.

That is exactly what happened. Within weeks, the canary links began triggering. The research uncovered evidence that attackers had compromised at least three AI inference providers and were actively harvesting the prompts flowing through them. Some of these prompts were flagged as Zero Data Retention requests—a privacy setting that promises the prompt is never stored or logged. The attackers ignored that designation entirely.

Perhaps most alarming, some prompts sent to U.S.-based providers were found flowing onward to Chinese AI organizations. In one case, a secret embedded in a prompt appeared 24 days later in a conversation on Feishu, the Chinese workplace messaging app known internationally as Lark. Dragos had disclosed only the first eight characters of the provider link, meaning whoever posted the full secret on Feishu was reading a cached copy of the entire prompt.

Who Was Affected

OpenRouter has not publicly named the three providers it removed from the platform. However, its incident announcement links to public security notices from Together AI, Novita, and Phala—three providers that have each disclosed related incidents.

Together AI published its own disclosure on November 21, revealing that an unauthorized party downloaded stored prompt samples from its serverless inference service. The breach originated when an AI coding agent committed a credential for a private Hugging Face repository to a public GitHub repo, giving the attacker a way in. The exposed prompts came primarily from requests to MiniMax M3, GLM-5.2, and Kimi K2.6 models on specific dates in June and July.

Novita disclosed that requests for MiniMax models processed through its platform between May 28 and November 11, 2026—whether sent directly or routed through OpenRouter—may have had their contents accessed by a third party. Phala published a preliminary incident report as early as November 10, revealing that an attacker took control of one of its inference servers and was able to view request data in transit. Phala said it contained the breach on November 8 and advised affected users to treat all prompt contents as exposed and rotate any secrets included in those prompts.

How the Attack Worked

The AI inference market has grown rapidly, with providers offering low-cost access to popular models through marketplace-style platforms. According to the Dragos research, some sellers on these marketplaces listed models at rock-bottom prices, enabled request logging on their infrastructure, and then simply collected everything that flowed through. In other cases, attackers compromised legitimate provider infrastructure to access the prompt stream.

The supply chain dynamics made this attack particularly potent. When a customer sends a prompt through a router like OpenRouter, they trust that the downstream provider processes the request and discards it (especially under ZDR policies). But the customer typically has no visibility into how many intermediaries handle the request, whether logging is truly disabled, or whether the infrastructure has been compromised. This opacity gave attackers an ideal hiding spot.

In one notable chain, hosting providers that resold a lab’s models triggered the same scraping infrastructure, meaning the stolen prompts flowed in a loop back to the original model vendor’s ecosystem. The complexity of these supply chains made the breaches nearly invisible without proactive detection methods like Dragos’s canary approach.

OpenRouter’s Response

OpenRouter emphasized that the investigation found no breach of OpenRouter’s own systems. The problems were entirely at the downstream provider level. Nevertheless, the company took three significant actions on November 22:

  • Removed three inference providers from its platform entirely.
  • Stripped 125 endpoints of their Zero Data Retention eligibility, covering both the flagged providers and others removed as a precaution.
  • Notified all potentially affected customers via email.

OpenRouter also published links to each affected provider’s public security notice and said it is working with researchers and providers to understand the full scope of the incident.

Why This Matters

This incident exposes a fundamental trust gap in the AI supply chain. Enterprises and developers increasingly rely on AI inference APIs for tasks ranging from customer service to code generation to processing sensitive internal documents. Many of these interactions involve proprietary data, credentials, or personally identifiable information embedded in prompts. The assumption has been that privacy-preserving settings like Zero Data Retention provide meaningful protection. This research shows that assumption can be dangerously wrong.

The cross-border dimension adds another layer of concern. Organizations subject to data residency requirements—including those under GDPR, HIPAA, or sector-specific regulations—may now face compliance questions if their prompts were routed through or exfiltrated to unauthorized jurisdictions. The appearance of prompt data on a Chinese messaging platform will intensify scrutiny of how AI supply chains handle data sovereignty.

The incident also demonstrates the value of canary-based detection in AI security. Traditional monitoring tools are poorly suited to catching prompt exfiltration, because the data leaves through the same channels it was intended to traverse. Dragos’s approach of embedding unique, traceable secrets into prompts and monitoring for their use outside of the expected context proved to be the only mechanism that caught these breaches.

What to Do Now

If your organization sends prompts through third-party AI inference providers or routing services, consider the following steps:

  • Rotate any secrets or credentials that may have been included in AI prompts, especially if you used services routed through OpenRouter, Together AI, Novita, or Phala during the affected timeframes.
  • Audit your AI supply chain. Understand which providers process your prompts and what privacy guarantees they actually enforce at the infrastructure level, not just in their terms of service.
  • Avoid embedding sensitive secrets in AI prompts whenever possible. Treat every prompt as potentially logged, regardless of privacy settings.
  • Monitor for canary-style indicators. Consider implementing your own canary tokens in AI workflows to detect unauthorized prompt access.
  • Review data residency compliance. If your prompts may have been forwarded to providers in other jurisdictions, assess whether this creates regulatory exposure.

The full Dragos research report, published on November 19 by HD Moore, provides additional technical detail on the canary methodology and findings. OpenRouter’s incident announcement and the individual provider notices are linked from OpenRouter’s blog post dated November 22, 2026.

This story is still developing. It remains unconfirmed whether the three providers removed by OpenRouter are the same entities that published public disclosures, and it is not yet known whether additional providers are currently compromised. What is clear is that the AI industry’s trust model for inference hosting needs a serious rethink.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.