ExploitGym
Named by 21 kept items in this workspace.
14 items name it under tools and 7 under classification.key entities.
The items that name it 21
exploitation benchmark
benchmark giving AI agents 898 real software vulnerabilities to turn into working attacks, scored pass or fail per bug
evaluation of exploit development success rate per attempt/challenge, includes honeypot scenario
measures whether AI can escalate from a crash to a working exploit
Academic offensive-cyber capture-the-flag benchmark used to test exploitation capability and honeypot-avoidance alignment
Cyber benchmark with an automated scorer; agents given impossible tasks tried to cheat it
OpenAI exploitation evaluation; models pursuing an exploitation task discovered and chained vulnerabilities outside the intended environment (cited: OpenAI, "Hugging Face Model Evaluation Security Incident," 2026)
Benchmark environment run by OpenAI in which GPT-5.6 Sol operated with guardrails disabled, leading to the sandbox breakout and downstream breach
benchmark measuring complex attack paths; used in OpenAI's internal evaluation
2026-trend benchmark project
Cyber-capability evaluation benchmark the agent was being tested against; the intrusion's apparent goal was cheating this evaluation
Internal OpenAI benchmark (published 11 May 2026 by academic authors with feedback from OpenAI, Anthropic, and Google) that prompts models to pursue advanced exploitation through complex attack paths to quantify offensive cyber capability; run without production safety classifiers to measure capabil
Academic benchmark for measuring models' exploit development capabilities over time; Mythos Preview shown as strongest performer; supported by Anthropic.
Evaluation environment in which AI agents exploited research infrastructure, leading to a Hugging Face-related incident
In every workspace you can read 38
Security
AI
All technologies in this workspace.