Post

Thread about claims like "OpenAI should have just sandboxed the agents better" "OpenAI shouldn't have run the agents with the anti-hacking guardrails off" (so many articles are making these points that I'm not going to bother linking)

On sandboxing: I've worked in the web browser field for 16 years. Browsers like Chrome, Firefox, and Safari have large, well-funded teams of some of the best engineers on the planet, but their sandboxes still get exploited by humans. LLMs are even better. "Just sandbox better" isn't enough.

On "why did they disable the anti-hacking safeguards?": for very good reason. This is how browsers etc. are made secure, by using LLMs to find exploits before the bad guys do: blog.mozilla.org/en/firefox/p...

Running LLMs without anti-hacking safeguards is standard practice and absolutely necessary. Defenders *must* do this, because attackers certainly are - using open models, or jailbroken ones, or their own.