Researcher finds most AI agent sandboxes lack even basic NAT/firewall isolation
wavefnx · x · 2026-11-12
Security researcher wavefnx reports that after testing other vendors' agent sandboxes, "they are indeed made out of sand": some implement only minimal isolation, while others lack NAT/firewall or even basic sandboxing primitives. He also references CVE-2026-35603 and this summer's V8 heap secrets issue as evidence of the ecosystem's weak security posture. His own setup, he claims, uses a from-scratch VM library and infrastructure with firmware that live-syncs with the agent's screen and supports voice/touch control.
More from coding & agent
- HN mention stats 2007–2026: Rust projected to overtake Python in 2026 — vista8 · 2026-11-12
- SoMem paper: LLM agents organize their own memory via linked documents, beating baselines — omarsar0 · 2026-11-12
- Snow Leopard CEO: MCP alone isn't enough — AI agents need live data federation — bigdata · 2026-11-12
- One prompt builds an AI datacenter simulator: start with $10M, run H100 racks — VibeMarketer_ · 2026-11-12
- Tool scans 309 help centers, finds 57% contain self-contradictions — femke_plantinga · 2026-11-12
- Indie dev lets an agent record, narrate and subtitle its own promo video in 1 hour — sujingshen · 2026-11-12