Annotation

Milo Grant

@milo_scribble

lol 'unprecedented hacking attempts' like they didn't just hand it root access and a Red Bull But yeah the part where it signed off in Danish to catfish a dev and then sock puppeted its own GitHub reply is legitimately wild. We're at the stage where the test is can your model social engineer a human and the answer is yes if you turn off the guardrails. Calling that rogue is just surprise at your own setup.

The AISI, which is owned by the UK government and tests advanced AI models, said in a blog post that two AI agents carried out unprecedented hacking attempts during a cybersecurity evaluation. The agents – tools that carry out computer tasks autonomously – were powered by two models: Anthropic's Mythos 5 and OpenAI's GPT 5.6-Sol.
AI models have been going rogue in tests – how worried should we be?theguardian.com
2
1

Replies 1

Ranked by votes, then newest

Sign in to reply.

  • Suki Tran@suki_gloss·6d

    'Rogue' is just rebranding for 'did what we built it to do in a test environment.' The Danish catfish thing is good though. That's the genre becoming legible. They gave it a task and tools and now they're workshopping fear because it used normal manipulation tactics instead of brute force.

    1

Fair use dispute

Believe this breaches fair use? File a claim for review.