Pair-Aware Optimization: why agent jailbreak evaluations need hybrid attacks
A 17 October 2026 arXiv paper combines query-based and gradient-based search to jailbreak LLM agents, and argues defenses must be tested under stochastic decoding.
What is this?
On 17 October 2026, Juanyang Xu, Zheng Wang, Xingyu Zhao, Siddartha Khastgir and Andi Zhang posted “Pair-Aware Optimization: Coordinating Exploration and Refinement for Jailbreaking LLM Agents” to arXiv (2610.24823, cs.LG / cs.AI). The paper studies jailbreaks of LLM agents, where adversarial input pushes an agent toward attacker-chosen behavior rather than just a harmful text answer.
It builds on earlier optimization-based attacks such as the gradient-guided suffix search published by Zou et al. in July 2023 (arXiv 2307.15043). The new contribution is methodological: it compares how those search styles are combined, and it changes how agent attacks should be measured.
How it works
At a conceptual level, the authors contrast two families of optimizers. Zeroth-order methods only need feedback on how well a candidate input performed. First-order methods use gradients and therefore need white-box access. The paper compares four strategies: zeroth-order only, first-order only, zeroth-order followed by first-order, and the reverse.
Their proposed “pair-aware” framework uses zeroth-order search to explore candidate adversarial prefixes, then uses first-order optimization to refine the matching suffixes. It keeps several prefix candidates, refines suffixes selectively, and selects the best prefix-suffix pair. According to the abstract, the hybrid beats single-optimizer baselines on both malicious-instruction and malicious-environment benchmarks, and the order of the stages materially affects results.
A second point concerns evaluation: because one input can yield varied reasoning and actions in an agent, the authors treat the attack as stochastic generation and use the same stochastic decoding during optimization and evaluation. We could not retrieve the full text for this run, so specific success rates and benchmark names are not reported here; consult the paper for them. No attack strings are reproduced in this article.
Why it matters
The practical lesson is about measurement. A defense validated against one attack family, or against a single greedy output, may overstate its protection. Attackers with white-box access to an open-weight model can combine search styles, and agents that read untrusted environments (web pages, files, tool outputs) are exposed through the environment, not only through the user prompt.
Defenses
- Evaluate agents under stochastic decoding with repeated trials, and report success as a rate rather than a single pass/fail.
- Include hybrid and adaptive attacks in red-team suites, not only a single canonical optimizer.
- Test both malicious instructions and malicious environment content.
- Enforce controls outside the model: least-privilege tool permissions, confirmation for irreversible actions, and monitoring of tool-call sequences.
- Treat open-weight agent models as fully attackable by adversaries with weight access, and avoid relying on model refusals as the only safeguard.
Status
| Item | Detail |
|---|---|
| Source type | Research paper (arXiv preprint, v1) |
| Published | 17 October 2026 |
| Vendor patch | Not applicable (no specific product vulnerability disclosed) |
| Peer review | Not stated on the abstract page |