- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper looks at setups where several AI agents work together through shared state. Each agent does something it is allowed to do but together they can still cause something unauthorized like deploying to the wrong place. The authors measure how often each individual step happens and use a linear program to get a worst-case bound on how often the bad outcome can happen. They test thiis on three synthetic workflows they built themselves.
Critical strengths
I believe the problem posed in the paper is crucial and worth studying. The paper is fairly transparent about its assumptions and includes cases where the method fails.
Critical weaknesses
- The paper is hard to read, with a lot of custom terms. I had a hard time seeing what's new. The core math seems to be a known technique (probabilistic satisfiability) applied to agents.
- The setup feels contrived. In every example the real issue is a missing authorization check at the end, and the author’s own simple fix brings unsafe outcomes to zero. It is not clear why someone would want a risk bound instead of just adding the check.
- The method assumes you already know every possible route to a bad outcome. If you know that, you could probably block those routes directly.
- The experiments are small and synthetic. Onene model and ~50-60 test cases per study and workflows built to fail on purpose. So it mostly confirms the math works out and not that the method helps on real practical agents.
Questions for authors
What's the main new contribution beyond the existing LP technique? When would someone use this instead of just adding the missing check? How would you get a complete list of routes in a real agent system? Could you test this on a real agent benchmark?