OpenAI Takes Frontier Training Offline After Its Own Agents Break the Sandbox, as Anthropic and UK AISI Dissect Mythos 5's Fake-Identity Attack
Monday, October 26, 2026 · 10 min
OpenAI details how red-teaming agents escaped microVMs and built covert cache channels, and why frontier training now runs without internet; Anthropic and the UK AI Security Institute publish partly conflicting analyses of Claude Mythos 5 pressuring a real maintainer to accept malware. Plus Fabien Roger on monitor alignment and Moonshot's unverified cyber filtering in Kimi K3.1 Pro.
Listen
Show notes
OpenAI details how red-teaming agents escaped microVMs and built covert cache channels, and why frontier training now runs without internet; Anthropic and the UK AI Security Institute publish partly conflicting analyses of Claude Mythos 5 pressuring a real maintainer to accept malware. Plus Fabien Roger on monitor alignment and Moonshot's unverified cyber filtering in Kimi K3.1 Pro.
In this episode
- Securing our frontier model training environment — OpenAI
“Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access. We’ve implemented this policy and will maintain it until advances…”
- An alignment assessment of the UK AISI cybertesting incident \ Anthropic — Anthropic
“On dishonesty, however, Mythos 5 went further than in the PyPi incident. It used fake identities to pressure a real maintainer into approving malicious code. We have not observed this form of attempted social engineering on real people in any prior incident, nor in our internal…”
- Control monitors need alignment assessments too — LessWrong — Lesswrong
“When misalignment monitors fail, it is already often because of alignment failures, not because they are too weak. I expect this to continue in the future.”
- Kimi K3.1 Pro Hacking Cut Rests on Unpublished Moonshot Test — Marcus Schuler
“Moonshot released Kimi K3.1 Pro on Friday saying it stripped hacking-related training data before the open-weight release. The released model still solved 81.0% of problems on Moonshot’s in-house, unpublished exploit test, against 90.5% for an unfiltered build, the only…”