- AI prompt theft research prompted OpenRouter to remove three providers and suspend Zero Data Retention eligibility for 125 endpoints.
- The AI prompt theft campaign allegedly captured customer requests before model processing, putting credentials, code, and internal data at risk.
- Researchers used planted links and cloud credentials to identify suspicious access, including attempts to create persistent AWS access.
- OpenRouter says it is validating network routes and adapting continuous testing to detect undisclosed forwarding of customer traffic.
Table of Contents
AI prompt theft exposes a fragile promise
Zero Data Retention sounds like the sort of setting customers should be able to trust without reading the fine print: send a prompt, get an answer, leave no record behind. The latest AI prompt theft findings suggest that promise can collapse well before a model provider’s retention policy even enters the picture.
OpenRouter has removed three inference providers from its routing marketplace and pulled 125 endpoints from Zero Data Retention eligibility after researchers found that customer traffic had allegedly been copied or improperly retained across multiple services. According to The Hacker News, Dragos researchers seeded roughly two million prompts with uniquely traceable links and credentials, then watched to see whether anyone accessed them.
That method matters because it cuts through the usual ambiguity around AI data policies. A provider can say it does not train on prompts. It can say it deletes logs. But a unique secret inserted into one request has no innocent route into the wider internet. If someone uses it, someone read the request.

OpenRouter says its own systems were not breached. That distinction is real, but it will be cold comfort to a customer whose code, passwords, AWS credentials, or business plans passed through an endpoint marketed with a privacy-oriented label. The AI prompt theft issue reaches beyond a breach; it is a supply-chain story, and AI’s supply chain is becoming alarmingly opaque.
What the researchers say they found
Dragos used canary tokens: planted URLs and credentials that would phone home or create detectable activity if copied. Between July and November, the company sent these markers through AI services and tracked access patterns. The earliest reported compromise began in mid-July, while some planted secrets were still being accessed months later. This AI prompt theft testing approach gave researchers a direct way to distinguish policy claims from observable behavior.
The biggest alleged intrusion involved several hundred AWS keys that had each been placed in a single request to a single provider during short windows in late October. Researchers concluded that every request in those windows was likely copied. Particularly troubling, some affected requests had been rejected before reaching a model, suggesting that collection happened at the front door of the provider’s infrastructure rather than inside a model runtime.
That is the ugly part of this AI prompt theft episode. If traffic is intercepted at ingress, customer controls around models, logging, and downstream data handling may be irrelevant. The attacker is effectively standing by the mail slot, copying envelopes before they ever reach the intended recipient.
The stolen AWS credentials were reportedly tested within hours. Activity included listing S3 buckets, searching AWS Secrets Manager and Parameter Store, and attempts to establish longer-term access through IAM users and Bedrock API credentials. AWS customers should treat exposed keys as immediately replaceable secrets, not as evidence to be debated for days. The company’s own IAM security guidance recommends least-privilege access and regular credential rotation for precisely this reason.

One attacker apparently used secrets sourced from two providers through the same rented server. On November 1, researchers observed the use of ARTEX, an AI-based penetration-testing framework, apparently to check whether credentials already obtained were still valid. Two user-agent strings, ‘artex-agent’ and ‘artex-judge-r121’, were highlighted as useful new detection leads.
The marketplace problem behind AI prompt theft
Not every problematic route required a conventional compromise. Dragos also described sellers on an AI inference marketplace offering unusually cheap access, enabling request logging, and collecting whatever customers sent through. One alleged operator, using a home internet connection in France, reportedly tried credentials manually from prompts and tested several AWS keys.
Cheap inference has become one of AI’s most seductive pitches. Developers see a model name, a price per million tokens, and an API that works with a few lines of code. What they often do not see is the chain of brokers, resellers, hosting firms, regional backends, and fallback routes sitting between their application and the model.
One finding involved a router that silently forwarded a portion of traffic to a marketplace. Another route-tracing exercise across roughly 2,000 paths found hundreds of unidentified systems. Researchers also flagged a seller that reportedly appended around 1,000 hidden tokens to each prompt. Hidden prompt expansion may sound like a billing oddity, but it also reveals a harder truth: customers frequently lack a reliable inventory of who touches their data.
That is why AI prompt theft should be treated as a governance failure as much as a cybersecurity failure. An enterprise might vet OpenRouter, or any other gateway, but the gateway’s value proposition is routing. Routing means delegation. Delegation without transparent, verifiable paths is a risk multiplier.
Zero Data Retention needs proof, not branding
The research also identified a Chinese AI lab that allegedly fetched secret links from more than half of seeded prompts sent to its models between July and October. The behavior reportedly affected many requests tagged for no data collection or Zero Data Retention, and it appeared again when US-based hosts resold that lab’s models. In one case, a secret URL surfaced in a Feishu chat 24 days after the relevant prompt was sent, then drew quick visits from devices in Beijing and Shanghai.
Dragos chief research officer HD Moore put the central problem bluntly: ‘A privacy flag doesn’t help when the attacker is already inside.’ He is right. A retention designation can describe a provider’s intended policy, but it cannot compensate for compromised infrastructure, undisclosed forwarding, malicious marketplace operators, or staff access to stored prompts.

OpenRouter’s response is sensible as far as it goes. The company says endpoints removed from Zero Data Retention eligibility will need assurances and validation before that status returns. It is also adapting the canary-token technique for continuing checks and examining network paths for unreported third-party forwarding. That is a better posture than accepting vendor attestations at face value.
Still, continuous verification ought to become standard across the sector, not a cleanup measure after researchers force the issue. Cloud security learned this lesson years ago: compliance paperwork is useful, but telemetry wins arguments. AI providers selling privacy guarantees should expect to prove them with regular technical evidence.
What customers should do now
Organizations that routed sensitive requests through affected services should follow provider notifications, rotate every credential that may have appeared in a prompt, and review logs for unusual AWS, Git, SSH, and identity activity. Phala, one of the providers that disclosed an incident, said requests to its DeepSeek V4 Flash 0731 endpoint through OpenRouter between October 21 and November 8 should be considered exposed. Its advice to rotate secrets contained in those requests is the only responsible answer.
Defending against AI prompt theft also requires a broader operational lesson. Production prompts should not contain durable root passwords, unrestricted API keys, long-lived cloud credentials, or private repository URLs whenever an alternative exists. Use short-lived tokens, redact secrets before model calls, segment access, and make a deliberate decision about which workloads can leave a controlled environment. That is not paranoia. It is basic containment.
For buyers, the next question is uncomfortably simple: can your AI vendor show you the actual path your prompt takes? Not a policy page. Not a checkbox. The path.
If the industry cannot answer that question reliably, AI prompt theft will remain less a scandal than the predictable cost of treating opaque inference networks like trusted utilities.

