What Happened and Who Discovered It
Dragos, an industrial cybersecurity firm, revealed that attackers gained access to internal systems at three AI inference providers that operate as part of OpenRouter's provider network. The breach remained undetected for months. None of the three providers knew they had been compromised until Dragos researchers disclosed the attack. Some of the intercepted requests were flagged with Zero Data Retention settings, meaning users believed their prompts would never be stored. Those requests were extracted and stolen anyway.
How the Attack Worked
The mechanics of this breach reveal why API security is harder than most users assume. Attackers positioned themselves inside the provider infrastructure, not at the network edge. This placement gave them access to requests after they arrived but before encryption or deletion processes could take effect. They captured both the input prompts and the model outputs. Because the attackers were operating from within the trusted network perimeter, traditional detection methods that flag suspicious external traffic would not have caught them. The theft was systematic and patient, collecting data over an extended timeline rather than triggering alert thresholds that might have been calibrated to catch rapid extraction bursts.
Why Zero Data Retention Failed
Zero Data Retention is a feature intended to prevent API providers from logging or storing user requests. Users enable it when they are concerned about data retention policies or want stronger privacy guarantees. In this case, the setting was bypassed entirely because the attackers intercepted the data at a layer below where retention policies are enforced. The feature operates as a software control inside the provider's application, but it cannot defend against an attacker who has gained system-level or database-level access. This distinction matters: the presence of a privacy control in the UI does not guarantee that data is protected from internal compromise. Users enabled what they thought was a privacy feature, but it provided no defense against the specific attack vector that was actually being exploited.
What Happened to the Stolen Data
The prompts and responses collected by the attackers have not been made public in bulk as far as reported. That does not mean they are secure. Stolen API requests often contain proprietary information, business logic, customer data, or sensitive instructions embedded in the prompts. An AI model's output can reveal information about how that model was tuned, what data it was trained on, or what instructions the provider uses to guide it. In corporate settings, prompts can encode proprietary workflows or confidential business reasoning. An attacker with access to months of this data can extract patterns, reconstruct workflows, or identify high-value targets for follow-up attacks. The value of the theft lies not in immediate resale but in the intelligence it provides for further operations.
OpenRouter's Response and Provider Verification
OpenRouter, which acts as a routing layer connecting users to various inference providers, removed the three compromised providers from its network after Dragos disclosed the attack. Users who had been routing requests through those providers would have been switched to alternative backends or informed that those providers were no longer available. The incident raises the question of how users can verify that a provider is actually secure. OpenRouter does not publish detailed security audits for each provider in its pool. The situation highlights a structural problem: users trust an intermediary (OpenRouter) to vet providers, but that trust is only as strong as the detection and response capability of those providers themselves. A provider that is breached but unaware of it will pass that vulnerability along to every customer using it through any gateway.
Reality Layer: What This Reveals About API Security
According to incident reports from industrial cybersecurity firms, insider or internal-network attacks on cloud inference providers typically go undetected for weeks to months because detection is tuned to flag external attack patterns, not internal data access. This matters because it exposes a gap between security theater (privacy settings, data retention policies) and security reality (access controls at multiple layers). Cloud service providers have published guidance that distinguishes between encryption in transit and encryption at rest, but less public material addresses the threat of compromise at the application layer where data is processed and forwarded. Users who select privacy options like Zero Data Retention are making a rational choice to protect their data, but they are not informed that this protection does not extend to every possible attack surface. The standard advice from security vendors is to assume that anything sent to a third-party API could be logged or accessed, regardless of what the provider's policy says. This attack proves that assumption correct.
Practical Steps for Users of AI APIs
If you use AI inference APIs or gateways like OpenRouter, consider the following:
- Review what data you are sending in prompts. Remove or redact personally identifiable information, business secrets, or sensitive context if possible.
- Do not rely on privacy settings alone as your only protection. Assume that any request could be logged or intercepted at some layer of the infrastructure.
- If your organization uses an API gateway or routing service, ask the provider or your vendor what security audit results are available. A provider that cannot or will not share security documentation is a warning sign.
- Check whether OpenRouter publishes a list of providers that have been removed or marked as compromised. Verify that your requests are routing through known-safe providers.
- If you work for a provider yourself, use this incident as a signal to audit your internal access controls and logging. Assume that data theft inside your infrastructure is possible and implement detection mechanisms that do not rely on external traffic patterns.
The core takeaway is this: a privacy feature or a data retention policy is not a guarantee. It is a control that works only if the entire infrastructure below it is trustworthy. When that infrastructure is compromised, the policy becomes irrelevant. The responsibility for verifying that infrastructure does not lie only with the provider; it also lies with users and organizations that choose which providers to trust.
Start today by listing the APIs and gateways your organization uses, then contact your vendor or service provider to ask what security audits or certifications they have completed. Request a summary of their internal access controls and detection capabilities. If they cannot or will not answer, that is itself information.
Source: The Hacker News
