escaping
5 stories and discussions about escaping, aggregated from every source we track.
1.
An AI sandbox is a sealed-off computer where agents run code and tools. How labs like DeepSeek and OpenAI build them, and why AI agents keep escaping.
2.
one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access
3.
The Yorba team was built to build. So our product failed to communicate "less." Norman and Conway showed us the fix isn't a feature - it's restraint.
4.
5.
Two ways out. One lets a patch write anywhere on the disk with no prompt. The other gets unsandboxed command execution out of read-only, the strictest mode Codex has.