logoalt Hacker News

user43928today at 8:25 AM1 replyview on HN

Alignment of the latest models is questionable, yes. That's a different topic.

For this particular gym incident, supposedly Opus 4.6 was used in OpenClaw, predating the current safety guardrails of Fable and co.


Replies

shaknatoday at 10:10 AM

Not really a different topic. All commands go into the same prompt system. If one part can accidentally be breached, then it can also deliberately be breached. Injection remains a problem.

show 1 reply