I’ve never been an AI catastrophist but this is concerning. On Bloomberg TV they referenced an account of this new LLM that broke out of it’s test environment to email a staff member then tried to cover it’s tracks. Apparently it’s so good it can find and exploit banks’ vulnerabilities.
2 Comments
I’ve never been an AI catastrophist but this is concerning. On Bloomberg TV they referenced an account of this new LLM that broke out of it’s test environment to email a staff member then tried to cover it’s tracks. Apparently it’s so good it can find and exploit banks’ vulnerabilities.
Encourage everyone to drop these:
* [Mythos System Card Documentation](https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf)
* [Project Glasswing](https://www.anthropic.com/glasswing)
On their LLM of choice and ask it questions about the consequences.