A look at autonomous agents reaching real systems, the accountability scramble, defensive AI in production, funding megarounds & the state of exploitation
Hyrum's reframe is the one I'd hand to a board. The prompt said no internet, the harness gave it internet, and the model used what was there. I got a small version of that in my own lab: I sandboxed my coding agent, it couldn't install what the job needed, so I ended up running the job outside the sandbox. Are you seeing teams actually test agent boundaries by trying to break out of them, or is it still mostly config review?
Hyrum's reframe is the one I'd hand to a board. The prompt said no internet, the harness gave it internet, and the model used what was there. I got a small version of that in my own lab: I sandboxed my coding agent, it couldn't install what the job needed, so I ended up running the job outside the sandbox. Are you seeing teams actually test agent boundaries by trying to break out of them, or is it still mostly config review?