Cybersecurity Company Uncovers Months‑Long Prompt Theft Across Three AI Inference Providers, Leading OpenRouter to Cut Ties
This week, industrial cybersecurity company Dragos reported that three AI inference services had been unintentionally exposing user prompts for months, a flaw that the services themselves did not detect until outside researchers sounded the alarm. Consequently, OpenRouter—a hub that aggregates AI models—said it will cut ties with the three affected providers.
The term "prompt theft" describes the illicit collection of the text that users feed into generative AI models—commonly known as "prompts"—which may hold proprietary code, confidential business details, or personal information. Since numerous AI services handle these inputs on distant servers, the data passes through the provider’s infrastructure, opening avenues for insiders or compromised systems to siphon the content.
The probe started after independent security researchers detected irregular network traffic associated with the three inference hosts. They passed their observations to Dragos, which performed a more thorough forensic examination and verified that threat actors had retrieved and duplicated prompt payloads over a prolonged timeframe. The three providers had not reported any knowledge of the breach before the public revelation.
Among the stolen requests, a number carried a "Zero Data Retention" flag—a setting that pledges not to retain user inputs after they are processed. Finding these flagged entries in the exfiltrated logs indicates that the attackers circumvented or disregarded the intended privacy protections, prompting doubts about the dependability of such guarantees throughout the sector.
OpenRouter’s move to sever ties with the three providers mirrors mounting pressure on AI service aggregators to guarantee the security of the underlying models they expose. The platform said it will reroute traffic to vetted alternatives and is collaborating with its remaining partners to introduce tighter monitoring and audit procedures.
The incident highlights a wider weakness in the fast‑growing AI hosting market, where rapid deployment frequently outstrips the establishment of solid security safeguards. Earlier this year, comparable cases of model‑stealing and data leakage were reported at other cloud‑based AI vendors, spurring demands for standardized security certifications.
Going forward, Dragos advises AI inference providers to implement end‑to‑end encryption for prompt transmission, impose rigorous access controls on internal personnel, and conduct regular audits of logging systems to spot illicit extraction. Regulators in multiple jurisdictions are also starting to assess whether current data‑protection statutes sufficiently address AI‑generated content, a discussion that could influence compliance mandates for the industry in the coming months.
Comments (0)
Be the first to comment.
Join the discussion