Assess risks of representation hijacking beyond jailbreaks

Investigate whether in-context representation hijacking affects subtler behaviors in large language models, including biasing reasoning chains, interfering with tool use, or manipulating decision-making in high-stakes domains such as legal or medical applications, beyond explicit jailbreak scenarios.

Background

Beyond demonstrating jailbreak efficacy, the authors raise concerns that representation hijacking could have broader impacts on model behavior in consequential settings. They explicitly frame this as part of their open questions, highlighting the need to systematically assess risks in reasoning, tool-use, and decision-making contexts.

References

First, beyond jailbreak scenarios, representation hijacking may pose risks in subtler domains, such as biasing reasoning chains, interfering with tool use, or manipulating decision-making in high-stakes contexts (e.g., legal or medical). Second, our work focuses on the attack surface and does not yet evaluate specific mitigation strategies. These open questions serve as stepping stones toward a new research frontier: representation-level alignment and defense.

— In-Context Representation Hijacking (2512.03771 - Yona et al., 3 Dec 2025) in Section 6, Discussion, limitations, and future work