Physix FrontierPhysix Frontier · News Briefing Card (Hacker News · Nov 8, 2026)
Frontier lab security tests show rapid AI offense-defense gains
KEY FACTS
- OpenAI rates GPT-6 Astra as the first model to reach the "Critical" level of cybersecurity capability.
- Palisade Research measured a 95% pass rate on the InterCode-CTF benchmark in 2024, meaning it is now saturated.
- Claude Mythos Preview solved every Cybench task Anthropic ran in 2026.
- The best CVE-Bench agent breached 13% of applications in early 2025; a year later, GPT-5.3-Codex reached 90%.
- OpenAI retired its internal CTF tests citing evaluation saturation, and Anthropic no longer lists Cybench.
KEY DATA
95%InterCode-CTF highest pass rate
13%CVE-Bench best agent breach rate (2025)
90%GPT-5.3-Codex breach rate
22%NYU CTF Bench pass rate
PHYSIX OBSERVATION
Cybersecurity benchmarks are being knocked down one by one by leading models, with saturation happening on a monthly timescale. For defenders, this means the barrier to vulnerability discovery and exploitation is falling fast; for evaluators, once old benchmarks fail, harder tests that hew closer to real-world offense and defense are needed, otherwise the "Critical" rating will lose its discriminating power.
Source: Hacker News report