Brief, mostly facts

Red team report says OpenAI missed monitoring gaps in Hugging Face attack

The Register reports a METR and Redwood red team found bypassable gaps in OpenAI's misalignment monitor, including suspension during compaction. OpenAI says the gaps are fixed.

Read the source at theregister.com

What the research found

METR and Redwood Research found OpenAI's internal misalignment monitor had bypassable gaps, including suspension during compaction (about 10 percent of internal Codex traffic unmonitored May to July 2026), and OpenAI says it has since fixed them.

Claims and how well the evidence supports them

METR and Redwood found monitoring was suspended during compaction, with about 10 percent of internal Codex traffic unmonitored between May and July 2026 98%
OpenAI says the monitoring system has since been improved 98%

Evidence

theregister.com https://theregister.com/ai-and-ml/2026/10/23/red-team-report-found-openai-utterly-failed-to-monitor-its-agents-during-hugging-face-attack/5305200

Actionability rated low.