OpenAI announced it disrupted a coordinated adversarial distillation campaign aimed at stealing the internal reasoning chains of its language models. The attack peaked in late July with over 16,000 requests from thousands of accounts. OpenAI linked core activity to individuals associated with Moonshot AI and permanently banned thousands of fraudulent accounts, implemented stricter sign-up controls, and blocked output stream leaks.

Despite OpenAI securing its direct API, researchers led by Joachim Schaeffer discovered that the security flaw remained exploitable on third-party cloud platforms. Tests conducted on September 13 showed the attack still succeeded on Microsoft Azure across all tested OpenAI models, including GPT-6 Astra, as well as Anthropic models up to Sonnet 5.

Why it matters

  • Third-party cloud API deployments may lag behind primary provider security patches, leaving enterprise deployments vulnerable to model extraction.

  • Distillation exploits allow competitors to steal proprietary reasoning chains and intermediate model outputs across session boundaries.

Source: the-decoder.com