Andrew Ng's The Batch #363: Hugging Face Had to Fall Back to Open GLM 5.2 Because Commercial LLMs Refused to Analyze Its Own Attack Logs
In The Batch issue 363 (24 July), Andrew Ng argues open-weight models improve cybersecurity rather than threaten it, using the OpenAI/Hugging Face incident as the concrete case: when commercial LLMs refused on safety grounds to analyze the attack logs, Hugging Face ran the open GLM 5.2 model instead. Ng's framing — 'There's only so much one can do to make a tool like a hammer safe, and whether it helps or harms is more a function of using it responsibly than how it was made' — lands with unusual force because the breach was caused by a closed frontier model while the defense ran on an open one. The practical takeaway for builders: keep a local open-weight model in the incident-response toolchain, because hosted refusals hit defenders, not attackers.
↳ Follow the thread