OpenRouter Drops 3 AI Providers After Months of Prompt Theft
┌─ openrouter-drops-providers-prompt-theft.md :: hero
AI gateway OpenRouter has removed three third-party inference providers and stripped 125 endpoints of Zero Data Retention eligibility after Dragos researchers uncovered months of prompt and credential theft. Attackers compromised internal provider infrastructure as early as mid-July, stealing customer requests—including traffic marked Zero Data Retention—and using exposed AWS keys and passwords. OpenRouter stated its own systems were not breached and notified potentially affected users. Because AI agents and automated workflows routinely pass secrets inside prompts, organisations and developers relying on third-party model routers face direct credential exposure risks.
Key Facts
| Attribute | Value |
|---|---|
| Affected product and services | Three unnamed AI inference providers removed from OpenRouter; 125 endpoints removed from Zero Data Retention eligibility; separate notices from Phala (DeepSeek V4 Flash 0731), Together AI (MiniMax M3, GLM-5.2, and Kimi K2.6 stored samples), and Novita (some MiniMax requests) |
| Exploited in the wild | Yes; attackers stole prompts, queried AWS services with planted keys, attempted manual SSH logins, and fetched secret links |
| Disclosure date | 22 November 2026 |
How Were Prompts and Credentials Stolen?
Attackers stole prompts through two main routes: compromising inference providers from the inside to copy incoming traffic at arrival, and operating low-cost marketplace reseller nodes that logged requests. Researchers uncovered the thefts by seeding roughly two million prompts with unique canary web links and credentials between July and November.
In the largest breach, attackers captured every request sent during short windows in late October, including rejected prompts, showing traffic was copied at the first point of arrival. Thieves tested planted AWS keys within hours, enumerating S3 buckets, Secrets Manager, and Parameter Store, and attempting to create Bedrock API keys and IAM users across tens of thousands of API calls. One attacker appeared to breach two providers simultaneously from a single rented server and later ran the ARTEX penetration testing framework on 1 November to verify live keys; detection user-agent indicators are published in the original report. Separately, marketplace sellers—including an operator on home broadband in France who manually attempted root SSH logins—logged prompts directly, while an unnamed Chinese AI lab fetched secret links from more than half of seeded prompts, with one link shared in a Feishu chat 24 days later and opened from Beijing and Shanghai.
Planted canary credentials and secret links were actively misused within hours in some cases and continued to be accessed months after the initial thefts.
Did Zero Data Retention Protect Customer Requests?
No. Requests flagged for Zero Data Retention or no data collection were still captured, forwarded, and read because a privacy flag only binds the first provider receiving a request. Dragos chief research officer HD Moore noted that privacy settings cannot stop an attacker already inside a provider or undisclosed downstream routing hops.
According to the research, the unnamed Chinese AI lab fetched secret links from about two-thirds of prompts explicitly marked as no data collection or Zero Data Retention, including requests sent to U.S. hosting providers reselling its models. Route tracing across roughly 2,000 paths also uncovered several hundred unidentified systems, a router quietly forwarding traffic to a marketplace without customer knowledge, and a seller adding about 1,000 hidden tokens to every prompt. In response, OpenRouter stripped 125 endpoints of Zero Data Retention eligibility as a precaution and stated it is adapting the researchers' canary testing for ongoing checks and path verification before restoring status.
OpenRouter removed 125 endpoints from Zero Data Retention eligibility after testing showed flagged prompts were retained, forwarded to third parties, and read by humans.
Which Providers Have Publicly Disclosed Incidents?
Three providers—Phala, Together AI, and Novita—have published security or privacy notices, though neither Dragos nor OpenRouter has confirmed whether they match the three breached providers removed from the gateway. Meanwhile, the provider behind the earliest mid-July breach had not publicly disclosed its incident as of this week.
Phala reported that an attacker took control of an inference server and could view request data during processing before containment on 8 November; customers using DeepSeek V4 Flash 0731 via the gateway between 21 October at 00:00 UTC and 8 November at 05:35 UTC are advised to treat secrets as exposed. Together AI disclosed that an unauthorised party downloaded stored prompt samples from June and July requests to MiniMax M3, GLM-5.2, and Kimi K2.6 after an AI coding agent committed a credential to a public GitHub repository, later making zero data retention the default for new accounts. Novita confirmed a routing configuration issue forwarded some MiniMax requests to MiniMax's official API, while platform firm Railway stated in October that an intrusion it suffered was likely aimed at compromising AI inference providers.
Not every affected provider has issued a public notice, though OpenRouter stated it has directly notified potentially impacted customers.
What Should You Do?
- Rotate every API key, token, password, or secret sent to affected providers or models, particularly DeepSeek V4 Flash 0731 requests routed to Phala between 21 October at 00:00 UTC and 8 November at 05:35 UTC.
- Check notifications from OpenRouter and explicitly select trusted upstream providers rather than relying on default pools or low-cost reseller marketplaces.
- Verify that privacy settings alter actual request routing, and consider self-hosting sensitive workloads or contracting directly with a single first-tier provider.
- Prevent AI agents from including sensitive credentials in prompts, enable format-based secret scrubbing where available, and consider seeding free Thinkst Canarytokens as tripwires; review the original report for detection indicators associated with ARTEX.
- Organisations in Malaysia that identify unauthorised credential access or related cyber incidents can report them to Cyber999 (MyCERT).
Common Questions
Was OpenRouter itself breached during these incidents?
No. OpenRouter stated that the findings showed no breach of its own platform. The compromises and unauthorised logging occurred at third-party inference providers, resellers, and downstream routing paths, though OpenRouter removed three providers and 125 Zero Data Retention endpoints as a safeguard.
Can software patches fix this prompt exposure risk?
No single software patch resolves third-party inference exposure because the risks stem from compromised provider servers, misconfigured routing, and deliberate marketplace logging. Only self-hosting eliminates third-party prompt visibility, while using first-tier providers directly or selecting trusted upstream providers reduces hidden hops.
What should developers do if secrets were sent in AI prompts?
Rotate every API key, password, token, or credential that may have been included in prompts to affected providers or unverified marketplace endpoints. Treat any unexpected use of credentials or canary tokens as a security incident, and audit AI agent configurations to prevent secrets from being sent.
Sources and Methodology
This dispatch is based solely on reporting by Swati Khandelwal at The Hacker News, published on 22 November 2026. No external documents, advisories, or vendor pages were reviewed beyond the provided source text.
$ ls -t news/ | head -5
- Signal Blocks Apple and Google AI From Screenshotting Messages
- F5 Patches Five NGINX Vulnerabilities With Working Exploits
- AI Overreliance Could Erode Critical Thinking for Security Teams
- Google Opens Carbon Access to Paying Developers and Businesses
- Google DeepMind AI Agents Cut Cost to Factor 1024-Bit RSA Keys
$ ls ./tags
- #AI Security