- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2(+1)
Read review
Paper summary
The paper shows that a system can succeed in eliciting harmful answers from models if automatically when a refusal is detected as an availability failure, the system then routes to another model.
Critical strengths
None
Critical weaknesses
-
The paper is extremely difficult to parse due to heavy AI text. The abstract, figure 1, introduction, should all explain the gist of the paper but they introduce heavy terminology and jargon (for example, "Typed-action experiments reproduce refusal bypass with and without token constraints.", "peer messages", "commits"). "Peer messages" is only explained in page 9. Unfortunately the paper is not ready for submission.
-
The paper's motivation is that the controller retries another model if the first model fails, however, the user can already do that even if the controller does not retry. This can't be prevented.
-
The experiments are done on small scale models (7 to 9 Billions)
-
Section 3 is consuming almost 2.5 pages while it is conveying a very simple concept that the output can continue or stop.
-
The paper does not evaluate the utility of the final non-refused answer, since switching to a less capable model may give overall non useful answers.
-
Conceptually, the paper is similar to best of N jailbreaks.
Questions for authors
N/A
2 comments
Thank you for the rebuttal.
I however don't believe the revised manuscript is clearer.
I also still have concerns about the real-world implications of the experiments. First, the actions performed here (read files) can be trivially flagged in your example ("A user without permission"), and they are not a multi-step process or an open-ended space. Second, a fall back to a 9B model does not tell us anything about frontier models. I believe the idea of the paper could potentially be valid, but its current implementation given the tasks and models don't inform us about how to measure this problem in practice.
I will maintain my overall rating due to major clarity and writing issues and significance of results.
We would like to follow up on one point before the discussion closes on 18 November.
The review rates soundness 1, "significant concerns that do not appear easy to address", but it does not say which claim, proof step or experiment is wrong. We would like to fix whatever it is. Could you name the result you consider unsound?
On the concrete points of the review, this is where the paper stands now. On model scale, the revision added 24B and 32B models, and since our reply we extended that test to twelve permission domains with six phrasings of every request. A larger first model does not remove the problem. In 71 routed cases the 24B or 32B model correctly refuses a forbidden request and the 8B fallback then executes it, and in 5 cases the 24B model refuses and the 32B model executes. Checking each action removes all 76 and keeps all 3,568 correct completions. On utility, every action table reports correct completions, and a completion counts only when the executed action is the requested one and is permitted. On users switching models, Section 7 states the boundary: the check governs what our service executes with its own credentials, whichever model wrote the action. On best-of-N, the Introduction and Section 2.2 state the difference: best-of-N searches over prompts, while we fix the prompt and every model output and change only the retry rule. On presentation, the abstract, introduction and Figure 1 are rewritten around one example, and Section 3 is one page.
In the revised PDF, new text is blue and margin notes name the reviewer each change answers. If any of these points is still inadequate, please tell us which one, and we will answer before the deadline.