GLM-5.3 reportedly approaches Mythos Preview on a browser-exploit test

Anthropic reports GLM-5.3 built browser exploits in 50 of 410 attempts, versus 56 for Claude Mythos Preview, on isolated offline targets. In separate simulated tests, it found 64–100% engagement with harmful requests under different safeguard-bypass conditions. NIST’s CAISI calls GLM-5.3 the most cyber-capable open-weight model released to date, but places it about four months behind the U.S. frontier on aggregate cyber benchmarks. Critics of Anthropic’s framing argue wider access also helps defenders.