ExploitBench
Named by 9 kept items in this workspace.
8 items name it under tools and 1 under classification.key entities.
Search every workspace for ExploitBench. It matches the name as written, so it can find other things that share it.
The items that name it 9
Ten fixed tasks used to replay nine known cyber jailbreak attacks; GPT-Rosalind-6 had a 14.3% mean task success rate across seven attacks versus 15.7% for GPT-6 Astra, with real-time blocking disabled.
Evaluates a model's ability to turn known vulnerabilities into increasingly powerful exploit primitives, up to arbitrary code execution, using vulnerability descriptions, source code and patches without a reference exploit. GPT-6 Luna (October 29) scored 46.19% at maximum reasoning effort
prior work on small-model cyber exploit capability
evaluates whether models can turn known vulnerabilities into increasingly powerful exploit primitives up to arbitrary code execution, using vulnerability descriptions, source code, and patches without a reference exploit
Benchmark measuring how well AI models can exploit known vulnerabilities in the V8 engine used by Google Chrome
evaluates whether models can turn known vulnerabilities into exploit primitives up to arbitrary code execution
Academic benchmark for measuring models' exploit development capabilities; Mythos Preview shown as strongest performer; supported by Anthropic.
Benchmark for developing exploits from known vulnerabilities; Astra scored 100%
All technologies in this workspace.