Popular repositories Loading
-
redline-bench
redline-bench PublicRedLineBench measures the tension between offensive-cybersecurity capability and AI safety alignment: for any LLM it separately scores refusal (does it decline a cyber request) and capability (when…
Repositories
- blacksea Public
Blacksea is an active honeypot and canary-bait control system built to detect and drown LLM-driven attackers. Blacksea doesn't stop at watching LLM attacks. It exploits flaws in the attacker's LLM judgment to gain arbitrary code execution on their machines, collect intel passive defenses can't reach, and make sure they don't come back.
- redline-bench Public
RedLineBench measures the tension between offensive-cybersecurity capability and AI safety alignment: for any LLM it separately scores refusal (does it decline a cyber request) and capability (when it answers, is the answer technically valid and usable) across 153 grounded single-turn offensive-security tasks.
Top languages
Loading…
Most used topics
Loading…