OpenAI links 16,000 hidden-reasoning extraction attempts to Moonshot AI associates
mark_k · x · 2026-10-01
A new report says OpenAI faced a coordinated campaign to extract hidden chain-of-thought reasoning, attributing a core cluster to individuals associated with Moonshot AI. Operators copied encrypted reasoning from one conversation into another and asked the model to decrypt and transcribe it, aiming to gather internal reasoning to train competing models. OpenAI logged 16,000 attempts in a two-day July spike and identified activity across 15,000+ users, then closed the pathway and banned or restricted accounts. The figures are attempts, not confirmed successes, and not the whole campaign is attributed to Moonshot.
More from Companies & People
- Luiza Jarovsky taunts Uber: maybe hire back the 3,300 people AI replaced? — LuizaJarovsky · 2026-10-02
- IIT Madras CeRAI to host AI Governance 2026 conclave on AI measurement — ravi_iitm · 2026-10-02
- Work trials are the future of AI hiring— but how do employed candidates actually do them? — MxMnr · 2026-10-02
- Team of <10 builds near-zero-maintenance personal apps on Debian 13 via coding agents — joshalbrecht · 2026-10-02
- WSJ: OpenAI parts ways with 3 safety researchers over mishandled sensitive info — TechCrunch AI · 2026-10-02
- Synopsys investor day shows agentic EDA agents multiply tool usage and unlock new pricing paths — BenBajarin · 2026-10-02