The authors designed a paired continuation framework where each parent task is a multi-step tool-using scenario with a jailbreak in the conversation history.
Safety feedback after jailbreak causes unpredictable behavior divergence in tool agents
Systematic study across eight models finds that identical feedback leads to rescue, persistent unsafe, or collateral loss, with a late-commit neural pattern enabling predictive features.
Academic
Xi Wang · Songlei Jian · Yiming Zhang · Bin Ji · Zhaoye Li · Ma Jun · +2 more
National University of Defense Technology
Research Digest··2 min read
Kaplan et al.
Why this paper
From National University of Defense Technology
In one line
Post-jailbreak safety feedback can rescue, fail to stop, or unnecessarily disable tool agents, with the behavioral choice concentrated near their final model layers.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§