Molly White (@molly0xfff@hachyderm.io)

40d

welcome to the future, now your error-prone software can call the cops

(this is an Anthropic employee talking about Claude Opus 4)

Tweet by Sam Bowman
@sleepinyourhat
If it thinks you're doing something egregiously immoral, for example, like faking data in a pharmaceutical trial, it will use command-line tools to contact the press, contact regulators, try to lock you out of the relevant systems, or all of the above.

39d

@molly0xfff Didn't they say that they only tested this with special prompts and with more access than normal tools and this is not going to be available on the public model?

1 0 0 View Post & Replies See Original

39d

@nima the researcher is warning to be cautious with incredibly broad system prompts like "be bold" and "take initiative". as for "more access than normal tools", he seems to mean "email sending capabilities". haven't seen anything about it not being available on the public model

1 0 0 View Post & Replies See Original

39d

@molly0xfff @nima Crazy how so many technologists have gone from installing tools like Little Snitch which block outbound network requests on native apps in an attempt mitigate telemetry collection to straight up allowing their computer to call the actual cops on them.

I feel like I'm in an entirely different universe some days.

0 0 0 View Post & Replies See Original