Discussion about this post

User's avatar
Nan Deng's avatar

Karotte killing stray processes before grading is isolation of the grading step, but the harder failures in the post are authorization problems inside the sandbox: rewriting the tests, reading the answer key, ignoring instructions the grader isn't checking. A process boundary doesn't help if those files are still writable by the attempt. I'd want the grader and the answer key to be a principal the agent can't act as, not just a process that starts after the agent exits.

Leon Wildcard's avatar

cool. do they have a thesis on how this helps with alignment?

Ready for more?