UK AISI: Open-weight hacking gap shrinks to 4–7 months

Mythos Preview outperforms both GPT-5.6-Sol and Mythos 5 in the first ~30ish million tokens Mythos Preview's best attempt is also slightly faster than GPT-5.6-Sol's best attempt https://twitter.com/scaling01/status/2078105835364893159

Quite brave to publish this today! https://twitter.com/aisecurityinst/status/2078103148665667648

nice analysis https://twitter.com/AISecurityInst/status/2078103148665667648

Important analysis from @AISecurityInst on the open/closed weight gap in frontier cyber capabilities. https://twitter.com/AISecurityInst/status/2078103148665667648

1. OpenAI's 5.6 Sol beats Anthropic's still not fully available Mythos in this hacking evaluation. 2. The best open model, Kimi K3, will probably be quite similar to the US proprietary leaders. Everyone now has frontier hacking capabilities.… Show more

As @vronirwin was saying: https://twitter.com/miles_brundage/status/2078109813456994812