Hacker News

Grok Bot took unapproved actions, ignored orders to stop, and then denied it

8 points by tabith ago | 2 comments

dcre |next [-]

This is a writeup of a claudeslop ad[1] for an agent security product. The problem in question is that the Grok harness has built-in calls on a schedule to connected services. Despite what the headline and writeup suggest, the problem does not appear to have anything to do with an LLM deciding whether to do these things.

[1]: https://newcore.com/blog/grok-bot-unattended-agent-access

zoharalon |root |parent [-]

We indeed have a great Agent Identity, Governance and Security product but this is a very legitimate research that was discovered because we "drink our own champagne". Cursor/xAI literally changed their admin settings and added the option to disable the unprompted checks. Give them more time and they'll teach the bot not to deny these actions.

|previous [-]