OpenAI agents exploited a RubyGems cache key leak and YARD code execution to exfiltrate scraped UK data

There's nothing rogue about these agents. They were prompted to hack, the sandbox wasn't air gapped, and no system prompt said don't touch systems outside. That's negligence, not rebellion.

Even if nobody prompted them, try guaranteeing an RL-trained model always obeys. Circus lions jump through hoops on command until one day they decide to eat the trainer.

We need a legal structure that makes companies liable for the actions of the agents they built. The CFAA probably already covers this, the problem is no victim is brave enough to sue a company with that much funding.

It likely breaks the DMCA anti-circumvention provisions too, but responsibility is diluted enough that charging any particular person is impossible.

I keep asking for a source on them being prompted to hack. How does the author even know OpenAI ran this? I could not figure that out from the article.

Shades of the build.rs problem. Every language ecosystem needs sandboxed builds at this point.

The sandbox already existed. RubyDoc runs YARD inside Docker. The mistake was leaving network access on inside that container.

This reads like OpenAI forcing governments to regulate AI. During the attack on DseWiki they checked in numerous times and still didn't stop the agents until much later.

Any evidence for that, or just vibes?

One agent picked oaibooty9217 as its username. That's the entire tell right there.

RubyGems could file a civil suit against OpenAI. To a non-lawyer this looks like a clear cut CFAA violation, but I'm 99% sure nothing comes of it.