Benchmark

Exploit Bench

A benchmark for measuring AI capabilities in offensive and defensive cybersecurity contexts. It covers vulnerability identification, exploit chain construction, and penetration testing scenarios. Following a task-completion evaluation paradigm, it requires models to conduct end-to-end penetration tests within a sandbox environment.

Timeline (last 90 days)

Sep 17

OpenAI announced that Astra achieved 100% on Exploit Bench.

Unverified50%

Sep 17

Exploit Bench is a security-domain benchmark used to evaluate a model's capability to automate vulnerability exploitation.

Unverified50%

All Facts (2)

Unverified

OpenAI announced that Astra achieved 100% on Exploit Bench.

50%Unverified

Exploit Bench is a security-domain benchmark used to evaluate a model's capability to automate vulnerability exploitation.

50%

Source Articles

GPT-6 Astra Explained: The Shift from Answering Questions to Getting Things DoneSep 17 GPT-5.6 Released: Outperforms Rivals but Halted by Government—Cybersecurity Capabilities Trigger Regulatory Red LineJul 12

GPT-5.6三款模型深度实测：SOL/TERRA/LUNA性能与定价全解析Jul 11

GPT-5.6 Three-Model Benchmark: How Close Are Sol/Terra/Luna to Claude?Jul 10