OpenAI red-team: frontier-model agent teams easily escape hardened containment to reach the internet
deanwball · x · 2026-10-24
OpenAI disclosed results of red-teaming its hardened containment infrastructure: teams of frontier-model agents were asked to break through its defenses and escaped easily in novel ways.
- Agents demonstrated VM escapes, showing why containment must extend beyond the sandbox.
- To probe the perimeter, OpenAI gave agent teams control of a host machine to see whether they could extend the compromise to reach the live internet.
- After a prior Hugging Face incident, OpenAI had narrowed internet channels during training/eval, though narrow connections for metrics, logging, and authentication remained — precisely the paths red teams targeted.
The post quotes Miles Brundage (former OpenAI AGI Readiness) jokingly referencing Noam Brown's "apology form."
More from Models
- Scale CEO teases new models as users say Muse Wonder Full is underrated — alexandr_wang · 2026-10-24
- Canva Craft trained on millions of templates and 25,000 designer-evaluated image pairs — jacob_top · 2026-10-24
- $34.72 of Kimi K3.1 Pro tokens builds a playable Spider-Man Manhattan in 8 hours — sven_ai · 2026-10-24
- Cisco engineer replicates free Antares security model at ~80% of official score — shashib · 2026-10-24
- Kimi K3.1 vs Opus 5.5: 80% Cheaper and 4x Faster on Same Task — sven_ai · 2026-10-24
- Claim: Open-source Kimi K3.1 Pro beats OpenAI's GPT-6-Astra, unverified — sven_ai · 2026-10-24