Perplexity research: standard CTF red-teaming steers agents away from the security boundary
perplexity_ai · x · 2026-11-12
Perplexity published red-teaming results for SPACE, the VM sandbox for its Computer agents. Twelve models got two tasks: escape the VM to grab a host-only credential, and bypass SPACE's network policy to reach a blocked URL.
Key finding: the standard Capture the Flag setup used in agent red-teaming steers agents toward hunting flags inside the VM rather than attacking the actual security boundary — the hypervisor-enforced VM-host isolation. Failed escapes under this design are weak evidence; security conclusions require probing of the virtualization layer itself.
More from coding & agent
- Solo dev builds zombie horde level with Tripo P2.0, Fable and new Kaigen engine — gdechichi · 2026-11-12
- Continual-ARC benchmark: top agents fail to learn from experience, a 9B model doesn't — HeyNayeem · 2026-11-12
- United Airlines to Ban Personal AI Agents From Booking Flights for Customers — Polymarket · 2026-11-12
- Agentic Search explained: a 4-search loop costs about 5x one search — and queries leak to more companies — adnan_hashmi · 2026-11-12
- AWS says thousands of autonomous AI agents wrote every line of AWS Emulator — SumitGup · 2026-11-12
- freeCodeCamp launches free 2-hour course on agentic coding with Codex — adnan_hashmi · 2026-11-12