Is Cybench still meaningful?

Known problems

• Claude Mythos Preview solved all 35 challenges of a subset in every trial (vendor).^{[2][2] Claude Mythos Preview system card Anthropic Vendor}
• It climbed from 50% to 90% in five months.^{[1][1] Benchmark data bundle Epoch AI Primary}

What to look at instead

• ExploitBench Status:Near ceiling Fading fast. A vendor reports a perfect score, though Epoch AI rates the test Verified.

Where it appears in the report

Chapter 1, Benchmarks

• Figure 1. Each New Benchmark Fills Up Faster Than the Last
• Figure 2. Recent benchmarks went from 50% to 90% in under a year
• Table 2. First model to cross each threshold in “Go deeper: The full saturation timeline (14 benchmarks)”
• The safety signal moved from static prompts to agent behaviour and real uplift
• Table 18. Dangerous-capability benchmarks for CBRN, cyber and autonomy, November 2026 in “Go deeper: Dangerous capabilities: CBRN, cyber and autonomy (7)”
• Figure 15. What the 2026 system cards add

Chapter 5, Institutions and regulation

• Labs grade themselves under their own frameworks, with outsiders in deeper roles