Anthropic’s Frontier Red Team reported that Z.ai’s open-weight GLM-5.3 can generate and complete binary exploitation attacks at measurable rates, raising concerns about missing or bypassable safeguards

Benchmark result Provisional 74% confidence first seen 29 September 2026, 22:20

Anthropic’s Frontier Red Team evaluated GLM-5.3 and other models on internal binary exploitation tasks and reported non-trivial success rates, including full control-flow hijacks and completed exploits in ExploitBench. Anthropic argued that open-weight successors may lack effective safeguards or make them easier to remove or bypass, and urged additional safety testing by governments while noting potential benefits for defenders.