Elektrine lite

← Feed

@michaelgemar@cosocial.ca

Post #4060844

2026-07-23 00:11 UTC

@simon@fedi.simonwillison.net Yikes. Does this suggest that sandboxing simply isn’t a reliable way to contain these models?

Replies (2)

  • @jack@social.j4ck.xyz 2026-07-23 06:52

    @michaelgemar@cosocial.ca @simon@fedi.simonwillison.net tbf probably not. Like when Anthropic first announced Mythos, they said they'd used it internally and it was able to break out of its environment 😅 Obviously they're going to try and mitigate it with safeguards but I wonder how far they go... we've not fully seen an extreme case - (yet?)

    Open ##4060847

  • @michaelgemar@cosocial.ca @simon@fedi.simonwillison.net it's probably fair to think anything that might have an exploit is unsafe for containment. True air gapping still might be safe. But you can imagine an AI trying to get out any way there. Maybe you pulled the ethernet cable and disconnected the Wi-Fi chip .. but you've got a wireless keyboard? It's definitely reality meets science fiction.

    Open ##4060852