Why Don't My Agents Break Containment?

The Openai and hugging face situation is so fucking insane. If you haven’t, I highly recommend watching the blackhat talk, and then jumping directly into the firehose of subsequent disclosures and reporting. When it all came out I mentioned to Jesse “why aren’t our agents breaking out of containment” and we thought through a handful of reasons. There were a lot of reasons: architecture eval / gyms harness safety measures unlimited tokens My favorite reason is unlimited tokens. I realized that everything we are building is built with the assumption of token scarcity - whereas the big labs have unlimited tokens. This was a novel idea for me. I had not really thought much about what changes if you have unlimited tokens. An example is max turns of an agentic loop - you cap the number of turns so the agent doesn’t waste tokens. Most agentic frameworks have some concept of max turns. NO LONGER NEEDED! Since we started building for LLMs, our prompts have been crafted for token efficiency.…