- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper focuses on security risks arising from the persistence of authorization decisions in long-lived LLM agents. The authors introduce the notion of residual-authority replay, in which permissions legitimately granted during benign tasks remain effective and are subsequently exploited in a different, potentially malicious execution context. The paper considers the authorization mechanisms in eight production coding agents and proposes a methodology for identifying permissions required by a target attack, acquiring these permissions through benign interactions, and subsequently reusing them during adversarial execution. The evaluation includes 508 AgentDojo attack cases across six LLM families, as well as 55 Terminal-Bench cases involving three production coding agents. The results show that persistent authorizations can significantly increase the success rate of attacks that would otherwise encounter user-approval requirements.
Critical strengths
The paper addresses a practically relevant security problem in increasingly autonomous and long-lived AI agents. The comparative analysis of authorization mechanisms in existing agent systems is useful, as it highlights differences in permission granularity, persistence, and reuse. The paper also includes the results of a large experimental evaluation. The distinction among harmful-action attempts, authorization decisions, and successful execution helps clarify the specific contribution of retained permissions to attack success.
Critical weaknesses
-
The security problem addressed by the paper has very limited novelty. My primary concern is the limited conceptual novelty of the paper. The security problem investigated by the authors, that is, that permissions granted for a particular task or context may remain effective after that task or context has ended, is well established in access control and authorization research. Fundamental principles such as least privilege, context-sensitive authorization, limited permission lifetime (see for example the notion of TRBAC), and revocation have long addressed precisely these concerns. Permissions should be constrained not only in terms of the operations and resources they authorize, but also with respect to the conditions under which they may be exercised. The paper tries to present the residual-authority replay as a new attack category. However, the underlying mechanism appears to be a specific manifestation of a known authorization problem: permissions that are insufficiently restricted in scope or lifetime can subsequently be reused in unintended contexts. The paper acknowledges closely related research on authorization continuity, lingering authority, and task-scoped authorization. However, I'm not convinced that the distinction drawn between this prior research and residual-authority replay establishes a sufficiently substantial conceptual contribution.
-
The proposed attack methodology has limited technical novelty. The proposed methodology consists of identifying the permissions necessary for a target malicious operation, finding benign tasks through which those permissions can be legitimately acquired, and subsequently exploiting the accumulated permissions in a different execution context. Although the methodology is clearly presented, its technical novelty is not apparent. The backchaining process essentially maps required operations to previously acquired authorization grants. Given an authorization system in which permissions persist across tasks, it is reasonably foreseeable that an attacker could exploit benign interactions to accumulate the permissions necessary for subsequent malicious operations. Actually in real life attacks, one step often carried out by the attackers is to escalate privileges in various ways (see for example the well known 2014 Target attack, in which the attackers were able to stole a token used by a system administrator for single-sign on and then used the privileges of the administrator, to create their own system administrator account). So the attack approach proposed in the paper is one of the possible ways agents could get permissions. It is therefore unclear whether the proposed approach introduces a genuinely new attack-generation technique or simply provides a systematic experimental procedure for demonstrating the consequences of persistent permissions.
-
The authorization problems in agentic systems is narrowly addressed. Another important limitation is that the paper investigates only a relatively narrow dimension of authorization security in agentic systems. Authorization in agentic environments involves substantially more than the persistence of permissions across tasks. In particular, multi-agent systems introduce additional complexities involving permission delegation, transitive authorization, accumulation of privileges, and preservation of authorization constraints across agent interactions. For example, an agent may receive permissions for a particular task and subsequently delegate some of these permissions to other agents. Unless delegation is carefully constrained, downstream agents may acquire permissions without the original task scope, execution context, or delegation restrictions being preserved. Moreover, permissions obtained through different delegation chains or accumulated over multiple tasks may collectively provide an agent, or a cooperating group of agents, with significantly more authority than intended. These issues are closely related to the problem addressed by the paper but are not considered in its threat model or evaluation. I am not suggesting that the paper must address all aspects of multi-agent authorization. However, the authors should clarify that their contribution is narrow and limited, and avoid presenting permission persistence as a newly identified or comprehensive authorization problem for long-lived agents.
-
The experimental results largely confirm expected consequences of persistent permissions and are thus not surprising. The experimental evaluation is extensive, but the significance of its findings should be considered in relation to the novelty of the underlying problem. The experiments compare attacks under fresh authorization states with attacks under states containing previously acquired permissions. In the controlled evaluation, unauthorized security-sensitive operations are stopped at the approval boundary in the fresh state. When previously acquired permissions match those operations, the operations can execute without renewed approval. Consequently, the observed increases in attack success rates are largely consistent with the expected behavior of the authorization mechanisms being evaluated. The experiments quantify the magnitude of this effect, which is useful. However, quantitative confirmation of an established security weakness does not, by itself, establish a substantial new technical contribution. Furthermore, the evaluation of task-aware authorization mechanisms demonstrates that restricting permissions to the current task can substantially reduce or eliminate the tested replay attacks. This is consistent with established least-privilege and context-sensitive authorization principles.
Questions for authors
None