Smart Contracts

Quantifies what frontier agents can exploit autonomously. Exploit revenue in simulation roughly doubled every 1.3 months, and a GPT-5 run cost about as much as it earned, so defenders should assume this is already feasible.

• articleExploit Evals
• articleAttack Navigator
• articleN Days
• articleScaling Laws And Interpretability Of Learning From Repeated Data

• articleExploit EvalsAnthropic research
• articleOpenAI’s accidental cyberattack against Hugging Face is science fiction that happenedSimon Willison
• blogOpenAI accidentally hacked Hugging Face — should we have seen it coming?Alexander Barry

Checking sign-in…

Loading comments…