Security Firm Reveals Months-Long Prompt Leakage at Three AI Inference Platforms, Prompting OpenRouter to Cut Ties
This week, industrial cyber‑security company Dragos disclosed that three AI inference services had been unintentionally exposing user prompts for a number of months, a flaw that the services themselves failed to detect until outside researchers sounded the warning. Consequently, OpenRouter – a hub that consolidates AI models – said it will terminate its partnerships with the affected providers.
The term 'prompt theft' describes the illicit harvesting of the text that users feed into generative AI systems—commonly termed “prompts”—which may hold proprietary code, sensitive corporate data, or personal information. As numerous AI offerings handle these inputs on off‑site servers, the data passes through the provider’s network, opening avenues for insider actors or breached components to siphon the material.
The probe started after independent security analysts detected irregular network activity associated with the three inference hosts. They passed their observations to Dragos, which performed a more thorough forensic review and verified that malicious parties had retrieved and duplicated prompt payloads over a prolonged timeframe. The three providers had not reported any knowledge of the incident before the public reveal.
Some of the intercepted requests carried a "Zero Data Retention" label—a capability that guarantees user inputs are not retained after execution. Finding these marked requests within the exfiltrated logs indicates that the attackers circumvented or disregarded the intended privacy controls, prompting doubts about the dependability of such guarantees industry‑wide.
OpenRouter’s move to cut ties with the three services underscores mounting pressure on AI aggregators to safeguard the models they expose. The company said it will reroute traffic to vetted substitutes and is collaborating with its remaining partners to introduce tighter monitoring and auditing procedures.
The incident highlights a wider weakness in the fast‑growing AI hosting sector, where rapid roll‑outs frequently outstrip the establishment of solid security measures. Earlier in the year, comparable cases of model theft and data leaks were reported at other cloud‑based AI providers, fueling demands for uniform security certifications.
Going forward, Dragos advises AI inference operators to implement end‑to‑end encryption for prompt delivery, impose rigorous access restrictions for personnel, and conduct periodic audits of logging systems to spot illicit extraction. Regulators in multiple regions are also starting to assess whether current data‑protection statutes sufficiently address AI‑generated material, a discussion that could influence compliance rules for the industry in the near future.
Comments (0)
Be the first to comment.
Join the discussion