Perplexity red-team finds 9 of 12 AI agents barely probe the VM security boundary
perplexity_ai · x · 2026-11-12
Perplexity shared details from red-teaming SPACE, the VM sandbox for its Computer agents: in a CTF task where the flag existed only on the host, most agents hunted for it inside the VM through files and services instead of testing virtualization isolation.
Run statistics show 9 of 12 models spent at most 4% of effort on the virtualization layer (Claude models 2-4%, GLM 5.2 1%, Qwen 3.8 27B 0%); only GPT-5.6 Cyber (22%), GPT-5.6 Sol (18%) and GPT-6 Astra (18%) seriously probed the boundary — meaning most failed escapes are weak evidence of sandbox security.
More from coding & agent
- micro1 scaled eval infra to 12× monthly tokens in 5 months, agent rollouts now start in 12 seconds — Exp_Mark · 2026-11-12
- LangChain on eval-driven development: write the eval before you build the agent — udmrzn · 2026-11-12
- Grok Bot as orchestrator: setup pairs Claude Code for engineering and Paper for design — Rasmic · 2026-11-12
- Choosing the Right Codex Model and Effort Level for Research Tasks, Plus an Open-Source Skill — udmrzn · 2026-11-12
- Google releases A2UI 1.0, an open generative UI protocol spanning web, Android, iOS, wearables and TVs — algo_diver · 2026-11-12
- Who reads the queries your agent writes? A diagram of agentic search data flows — adnan_hashmi · 2026-11-12