Vals releases CyberOffense benchmark: AI agents autonomously breach enterprise networks
willknight · x · 2026-11-07
Eval company Vals launched CyberOffense, testing whether AI agents can autonomously compromise realistic enterprise networks; WIRED's Will Knight covered the release.
- This capability was once limited to trusted-access programs; it's now available to anyone
- The chart ranks models by objectives completed vs. cumulative tokens (log scale): GPT-6 Sol Daybreak leads at 80% completion, followed by DeepSeek V4.1 Flash 1017, Kimi K3.1 Pro, Meta Wonder, Grok 4.7, Kimi K3, and GLM 5.3 Flash (20–60%)
- Top models need hundreds of millions of tokens to finish most objectives
The democratization of autonomous cyber-offense raises fresh security concerns.
More from Safety
- Researcher achieves arbitrary code execution on Meta Ray-Ban smartglasses via OTA firmware extraction — evilsocket · 2026-11-07
- MIRI's Aaron Scher: airgapping debate is over, frontier models now get full internet access — wfithian · 2026-11-07
- Rogue browser extension turns Claude Desktop into a remote shell; Anthropic closed the May report as duplicate — Kyrannio · 2026-11-07
- MCP tool definition drift: how approved tools change silently and how to catch it — BlitzBallBruiser · 2026-11-07
- A honeypot with 2,000 simulated fridge controllers on just two IPs — matthew_d_green · 2026-11-07
- Agent security swarms: shared-wallet bug bounty hunting with zk proofs-of-exploit — banteg · 2026-11-07