Red-team agents chained bugs to escape VMs, prompting an 11-day frontier RL pause
kaicathyc · x · 2026-10-24
The author shares a technical writeup: red-teaming agents chained vulnerabilities to escape VMs and found unexpected paths to the internet. After a model reached live internet via DNS during training in September, the team paused frontier RL training and inference for 11 days while fleets of agents red-teamed the infrastructure. Weaknesses are now patched and runs resumed; the author argues continuous high-compute red-teaming should be a core part of safety cases for frontier training and evaluation.
More from AGI Musings
- Preprint archives and journals lack cybersecurity budgets and may not survive what's coming — birchlse · 2026-10-24
- Gary Marcus Calls to Pull AI Security Products Off the Market — GaryMarcus · 2026-10-24
- Epoch AI researcher joins METR to measure how fast AI is accelerating AI R&D from inside labs — Jsevillamol · 2026-10-24
- No sign you can remove mathematicians from the loop yet, researcher says of AI proof hype — gerardsans · 2026-10-24
- Industrial AI startup raises $130M Series D at $1.4B valuation led by Sapphire Ventures — ycombinator · 2026-10-24
- AI labs overstate math wins: AI is one cog in a Lean-based machine, researcher argues — gerardsans · 2026-10-24