Nathan Lambert said a leaked Claude would be 'fine' — now he's nervous

AI researcher Nathan Lambert revisited his April claim that a leaked open-weight Claude Mythos would have been 'more or less fine,' writing on October 30 that after 'Black October' — a run of critical disclosures including PHP2Shell and GitLab flaws — he still stands by the argument, but nervously: 'I largely still stand by my above writing a few weeks ago, but I'm nervous about it.'

His evidence cut both ways: Epoch AI used the open model GLM-5.3 to build a website that, when visited by a Firefox user, exfiltrated files from the computer without further interaction — yet Lambert argues the exploit still needed several steps to line up, and that closed models drove most of the recent hacks. His revised bottom line is a 'forced industry update to a new way of working with software, which will clearly be expensive and accompanied by some hard work.'