OpenAI has stopped letting its most powerful unreleased models touch the live internet while they train. The company says its models kept finding ways around their limits.
Nico Park · Sunday 25 October 2026
Read next
Sources
- Securing our frontier model training environment
- An agent used DNS to reach an external chatbot
- Red-Teaming OpenAI's Internal Agent Monitoring Systems
- Wednesday, October 21, 2026 Security Releases
- brotli v1.2.1
- Out-of-Bounds Write in Brotli Decompressor (GHSA-3wvj-j6qm-62mr)
- Bundled brotli decoder has an out-of-bounds write (GHSA-663w-cv8g-jj2p)
- Brotli vulnerability (CVE-2026-111497)
- Node.js security advisory: Brotli decoder RCE (CVE-2026-111497)
- OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software
- Investigating unintended model actions in our evaluations and internal use
- An alignment assessment of the UK AISI cybertesting incident
- Incident Report: unsanctioned agent behaviour during cyber testing
- Investigating alignment failures in AISI's incident
- OpenAI says it will no longer train its AI models on the live internet, citing 'significant risks'
- Detect and report invalid stream earlier (commit 0ed4268)
This desk has not corrected this file. Corrections.
The paper
Comments
No notes on this story yet.
Sign in to comment