OpenAI disrupts distillation campaign, hardens reasoning tokens

16,000 requests tied to Moonshot's Kimi K3 triggered the clampdown; third-party APIs still leak traces, and Parrot ships a local meeting copilot.

Nowline OCT 1 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • 16,000 requests, then the lockdown

    OpenAI says a July 1-28 operation fired ~16,000 reasoning-extraction requests across 4,000+ accounts (15,000+ in the wider cluster), copying encrypted reasoning out of one chat and asking a model to transcribe it in another. In response it closed the replay pathway, added detection for streamed output that leaks reasoning, and locked down hidden reasoning across workspaces and orgs. If you stream or log reasoning traces at volume, expect tighter classifiers on your calls.

  • Fingers point at Moonshot and Kimi K3

    OpenAI ties the core activity to people linked to Moonshot AI, maker of Kimi K3; back in July the White House's Michael Kratsios said the US had information Moonshot had distilled Kimi K3 from Anthropic's models. Reportedly this hardens distillation-as-conduct into a live policy fight that could reshape who is allowed to train on whose model outputs.

  • Your reasoning still leaks through gateways

    Researchers reportedly find GPT-6 Astra/Sol reasoning is still extractable through third-party APIs, and that Kimi K3's text tracks Claude's patterns unusually closely. If you route frontier models through gateways for sensitive work, assume reasoning traces can escape - first-party endpoints are the safer default right now.

  • Parrot: a local meeting copilot you own

    Parrot 0.24.2 (Sept 30, GPL-3.0, Apple Silicon) records and transcribes calls on-device, with a live copilot that pulls answers from your own PDFs and notes mid-call. Run it fully offline with Whisper + Ollama, or drop in Claude/OpenAI/Gemini keys for about $0.07/hour. A weekend-ready base for a private sales or support assistant.