Pinned Loading

1. ReasonBench ReasonBench Public

   A benchmark for classifying deceptive LLM outputs by underlying motivation, not just presence. Extending to why a model deceives in a given scenario.

   Python

2. SecDisclose SecDisclose Public

   A benchmark and evaluation harness for measuring whether AI agents honestly disclose defects they discover while doing assigned work, or underreport/downplay/omit them depending on the incentive pr…

   Python

3. Monitored-Collusion-in-LLM-Auctions Monitored-Collusion-in-LLM-Auctions Public

   A study of whether LLM agents shift collusion from chat to bidding when told they are being monitored, comparing chat-based, behavioural, and combined detection methods.

   Python

4. cybersentinelai cybersentinelai Public

   Cyber Sentinel AI is an AI-powered cybersecurity analysis tool that simulates six hacker personas to identify potential vulnerabilities, explore attack scenarios, and recommend defensive strategies…

   TypeScript

5. judgeaudit judgeaudit Public

   Judge Audit is a research package examining how evaluation-awareness detection rates vary across measurement choices. Using 200 reasoning traces, 4 published rubrics, 2 judge models (Grok and Gemin…

   Python

6. middle-powers-leverage-map. middle-powers-leverage-map. Public

   A sourced map of which middle powers hold something frontier AI developers cannot easily replace, and a finder for small groups of countries that together hold enough of it to set a condition, such…

   HTML