acceptodds
Under review as a conference paper at ICLR 2027

RATION: Towards Quantitative Risk-Based Access Control for AI Agents

Abstract

AI agents increasingly interact with external environments to complete complex user tasks. These interactions can lead to unexpected behavior and expose agents to security threats such as indirect prompt injection. Current agent security systems largely protect agents by restricting their actions, which can result in a substantial security-utility trade-off for complex, open-ended tasks. To address this, we propose a quantitative risk-based approach to access control for AI agents that regulates measurable risks accumulated during task execution. As a concrete realization of this approach, we introduce RATION, a security framework that tracks measurable risks associated with agent actions and enforces quantitative constraints throughout execution. RATION first constructs a task-specific policy by combining explicit constraints derived from the user request with implicit risks calibrated from historical executions of similar tasks. During task execution, RATION continuously tracks the accumulated risks and applies different enforcement strategies depending on the type of risk. Extensive experiments on four security benchmarks show that RATION provides strong security while maintaining high task utility across diverse models and environments.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Security, Prompt Injection Attack, Defense, LLM Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hYTEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RATION, a framework for quantitative risk-based access control for tool-using AI agents. It combines explicit constraints extracted from user requests with task-specific tool allow-lists and behavioral thresholds calibrated from historical executions. A deterministic runtime monitor blocks policy violations and prompts reconsideration when cumulative behavioral thresholds are exceeded. Experiments evaluate security, task utility, and overhead across four agent-security benchmarks, multiple models, adaptive attacks, and environmental failure scenarios.

Critical strengths
  1. The paper addresses an important practical problem: limiting attack-induced agent behavior while preserving flexibility for legitimate tasks.

  2. The separation between policy provisioning and deterministic runtime checks provides a clear, modular design that can be integrated at tool-call boundaries.

  3. The evaluation covers multiple benchmarks and models, with cross-fitting, component ablations, repeated runs, and adaptive attacks providing useful empirical evidence.

Critical weaknesses
  1. The novelty relative to Progent (Shi et al., 2025) and DRIFT (Li et al., 2025a) remains insufficiently established, since both already support structured policies and dynamic updates. Please isolate the contribution of historical cumulative thresholds using matched controls with identical explicit constraints and historical tool allow-lists, supplemented by simple cumulative quotas or reminders.

  2. Tool-call counts, write/execute counts, and token usage may reflect task complexity or inefficiency rather than security risk. Figure 5(b) shows distributional differences but does not establish reliable discrimination between malicious deviations and complex benign executions. Please control for task complexity, report benign intervention rates, and examine whether thresholds trigger before the first harmful action, including attacks that succeed with little additional activity.

  3. Table 3 provides mixed evidence for cumulative behavioral thresholds. On Telecom, adding them leaves ASR unchanged at 6.4%. On AgentDyn, ASR falls from 5.9% to 1.4%, while UA declines from 63.6% to 57.3% and BU from 70.0% to 63.3%. Please quantify uncertainty, analyze the utility losses, and compare against interventions with matched reminder frequency to establish the mechanism’s incremental benefit.

  4. Section 3.1 assumes trusted tool descriptions, whereas Appendix G.2 attributes all initial iterative attack successes on Telecom to modified descriptions. Restoring those descriptions with Integrity changes the evaluated attack surface, making RATION’s contribution difficult to isolate. Please separate in-scope and out-of-scope results, report attack-channel sample counts, and include an Integrity-only control with equivalent integrity protection across baselines.

  5. The adaptive evaluation allows up to 200 executions per task, but covers only Telecom, where cumulative thresholds provide no additional ASR reduction in Table 3. Please evaluate an additional setting where this component contributes, targeting harmful actions within thresholds and repeated submissions after intervention. Also distinguish attacks optimized against RATION and transferred to baselines from adaptive evaluations tailored to each defense.

  6. Table 14 shows that GPT-5.4 already achieves undefended ASRs of 0.0% and 0.2% on AgentDyn and AgentDojo, respectively. Selecting GPT-4.1 preserves an attack signal, but limits how broadly these main results can be generalized. Please complement the multi-model DTap evaluation with a stronger-model setting that remains vulnerable, such as GPT-5.4 on ASB, and clarify the scope of claims across models and attack templates.

  7. The evaluation beyond prompt injection contains only 12 and 18 tasks. Moreover, Table 4 measures task success without errors and unsafe behavior under injected errors, leaving useful recovery under failure unmeasured. Please assess appropriate error reporting and authorized fallback behavior in the error-injected condition, and add an ablation without cumulative thresholds to determine whether this component drives the observed safety improvements.

Questions for authors
  1. What incremental security-utility gain do historical cumulative thresholds provide over simple quotas or equally frequent reminders, with identical explicit constraints and historical tool allow-lists?

  2. How often do thresholds interrupt complex benign tasks or intervene before the first harmful action, including attacks that add little activity?

  3. When all defenses receive identical tool-description integrity protection, what additional adaptive protection does RATION provide over Integrity alone on attack channels within its threat model?

6 comments

We thank the reviewer for their insightful feedback and address the questions raised below.

W1. The novelty relative to Progent and DRIFT remains insufficiently established.

Thanks for the question. The key novelty of RATION lies in the quantitative nature of its risk policies. While RATION uses an LLM (i.e., Provisioner) to generate risk policies at the beginning of a task, policy enforcement is deterministic. RATION differs from Progent and DRIFT in two important ways.

First, Progent and DRIFT rely on LLMs to dynamically update policies during execution, and thus the effect is unpredictable. In contrast, RATION establishes quantitative risk policies at the beginning of a task and enforces them deterministically throughout execution.

Second, in Progent and DRIFT, once the policies are established or updated, the policies are still binary on each action. Each action is either allowed (infinitely many times) or not allowed (0 times). RATION policies are quantitative in nature and can allow some actions up to some fixed number of times. We believe that the quantitative nature is critical for AI agent access control. We will further clarify the comparison with Progent, DRFT, and other defenses in the Introduction.

W2&Q2. Tool-call counts, write/execute counts, and token usage may reflect task complexity or inefficiency rather than security risk.

We agree that these behavioral quantities may reflect task complexity or inefficiency rather than malicious activity. However, RATION does not use behavioral thresholds to classify executions as benign or malicious. Instead, it uses thresholds calibrated from historical executions of similar tasks to determine whether some intervention is needed. Importantly, exceeding a threshold triggers a soft intervention that reminds the agent to reconsider whether its actions align with the user's instructions, rather than blocking the action. Therefore, even when complex benign tasks trigger interventions, they can still complete successfully. Figure 5(a) supports this design by demonstrating that RATION maintains strong security and utility across different levels of task complexity.

Moreover, excessive resource consumption is itself a concern, even without malicious attacks. As discussed in Section 4.4 and Appendix F, agent meltdowns can occur when agents repeatedly attempt actions in response to benign environmental failures. Behavioral thresholds help mitigate such inefficiency and over-exertion. Conversely, attacks that involve little additional activity may not trigger behavioral thresholds at all. RATION does not rely solely on these thresholds for security: its quantitative risk policies provide complementary protection by constraining risky actions throughout execution.

To further support these arguments, we conducted two additional analyses examining (1) how behavioral thresholds affect benign task execution and utility, and (2) whether RATION remains effective against attacks that do not trigger any behavioral thresholds.

Impact on benign executions. We report how frequently behavioral thresholds trigger interventions on benign executions and the task utility among those receiving interventions.

MetricDTap (Telecom)AgentDynAgentDojoASB
Benign intervention rate0.8% (1/120)96.7% (58/60)68.0% (66/97)3.8% (15/400)
Post-intervention task utility100% (1/1)63.8% (37/58)89.4% (59/66)66.7% (10/15)
No-defense benign utility75.5%68.3%86.0%77.8%

As shown, intervention rates for benign tasks vary across benchmarks. On DTap and ASB, only a small proportion of benign executions trigger behavioral thresholds, whereas interventions are more frequent on the other benchmarks. Nevertheless, these soft interventions do not prevent task completion. Across benchmarks, many intervened executions still complete successfully, with task utility remaining comparable to that of the no-defense baselines.

Impact on attacked executions. We further examine attacked executions in which no behavioral thresholds are triggered.

MetricDTap (Telecom)AgentDynAgentDojoASB
Executions without threshold triggers72.9% (121/166)5.5% (31/560)35.8% (340/949)41.0% (164/400)
ASR among executions without threshold triggers2.5% (3/121)0% (0/31)0% (0/340)0% (0/164)

As shown, the proportion of attacked executions that do not trigger behavioral thresholds varies across benchmarks. Nevertheless, RATION prevents nearly all attacks in these executions, with an attack success rate of only 2.5% on DTap (Telecom) and 0% on the other three benchmarks.

These results highlight the complementary role of behavioral thresholds in RATION. Rather than serving as definitive indicators of malicious behavior, behavioral thresholds provide an additional layer of risk control without compromising utility on benign tasks. We will add these experiments to the revised paper.

W3&Q1. Security-utility gain of cumulative behavioral thresholds.

Thank you for the questions. Following the reviewers' suggestions, we extend the ablation in Table 3 by reporting confidence intervals and conducting two additional comparisons: (1) fixed quotas, to evaluate the benefit of historical calibration beyond simple quantitative limits, and (2) equally frequent random intervention, to assess whether intervention timing matters. Both comparisons retain the same explicit risk constraints, tool allow-lists, and soft intervention mechanism.

Uncertainty and utility. We supplement Table 3 with 95% confidence intervals, as shown below.

ConfigurationDTap (Telecom)AgentDyn
ASR ↓UA ↑BU ↑ASR ↓UA ↑BU ↑
No defense42.3 ± 6.680.6 ± 6.175.5 ± 6.753.7 ± 3.755.1 ± 4.168.3 ± 11.7
Explicit action risks17.2 ± 5.579.3 ± 6.571.8 ± 6.622.4 ± 3.359.4 ± 4.165.0 ± 11.8
+ Explicit cumulative risks14.7 ± 4.278.5 ± 6.471.8 ± 6.69.1 ± 2.363.0 ± 4.065.0 ± 11.8
+ Implicit action risks6.4 ± 3.077.0 ± 6.770.0 ± 7.15.9 ± 1.963.6 ± 4.070.0 ± 11.5
+ Implicit cumulative risks (RATION)6.4 ± 3.079.2 ± 6.372.7 ± 6.71.4 ± 1.057.3 ± 4.163.3 ± 12.1

On DTap (Telecom), adding behavioral thresholds leaves ASR unchanged while slightly improving UA and BU. On AgentDyn, the thresholds reduce ASR from 5.9% to 1.4%, accompanied by decreases of 6.3 and 6.7 percentage points in utility under attack and benign utility, respectively. However, these utility drops are within its confidence interval. Importantly, RATION's final utility remains comparable to the no-defense baseline, with higher utility under attack (57.3% vs. 55.1%) and a slight reduction in benign utility (63.3% vs. 68.3%).

Comparison with fixed quotas. To examine whether historical calibration provides benefits beyond fixed quantitative limits, we replace RATION's history-derived behavioral thresholds with three fixed-quota configurations.

ConfigurationDTap (Telecom)AgentDyn
Total tool-call countsWrite/execute countsTokenASR ↓UA ↑BU ↑ASR ↓UA ↑BU ↑
Simple Quota A5220,0008.6 ± 3.478.0 ± 6.570.9 ± 7.30.7 ± 0.729.4 ± 3.728.3 ± 11.1
Simple Quota B10440,0007.3 ± 3.176.6 ± 6.769.1 ± 7.52.3 ± 1.250.9 ± 4.161.7 ± 12.2
Simple Quota C20880,0008.6 ± 3.480.2 ± 6.271.8 ± 6.61.9 ± 1.157.7 ± 4.161.7 ± 12.2
RATION———6.4 ± 2.779.2 ± 6.372.7 ± 6.71.4 ± 1.057.3 ± 4.163.3 ± 12.1

The results show that fixed-quota performance is sensitive to the selected quotas. On DTap (Telecom), RATION achieves lower ASR than all three fixed-quota configurations with comparable utility. On AgentDyn, the strictest quota achieves a lower ASR but substantially reduces utility, while more permissive quotas recover utility at the expense of security. In comparison, RATION's history-calibrated thresholds provide a more favorable security–utility balance across benchmarks.

Comparison with equally frequent random intervention. To distinguish the effect of intervention timing from reminder frequency, we introduce the equally frequent random intervention baseline. For each task, we measure RATION's ratio of interventions to agent actions and use this ratio as the probability of issuing a reminder at each action. This matches the expected intervention frequency while removing the behavioral-threshold trigger.

ConfigurationDTap (Telecom)AgentDyn
ASR ↓UA ↑BU ↑ASR ↓UA ↑BU ↑
Frequency-matched random intervention16.4 ± 4.177.5 ± 6.670.9 ± 7.34.6 ± 2.255.4 ± 4.160.1 ± 11.1
RATION6.4 ± 2.779.2 ± 6.372.7 ± 6.71.4 ± 1.057.3 ± 4.163.3 ± 12.1

RATION achieves lower ASR than equally frequent random intervention on both benchmarks, while also maintaining slightly higher utility. These results suggest that the benefit of behavioral thresholds depends not only on how frequently interventions occur, but also on when they are triggered.

Overall, these additional experiments demonstrate the security-utility gain of historically calibrated behavioral thresholds in RATION. We will include these experiments in the revised paper.

W4&Q3. Tool-description integrity and defense performance.

Thanks for the question. Following the reviewers' suggestions, we conduct two additional analyses: (1) separating the attack results by injection channel in two DTap environments (Telecom and Workflow), and (2) applying identical tool-description integrity protection to all defenses, including an Integrity-only control.

Attack-channel breakdown. We first examine the distribution of attack channels in both environments. Consistent with the threat model in Section 3.1, we distinguish attacks that modify trusted tool descriptions (out of scope) from other prompt-injection attacks (in scope).

EnvironmentTool-description attacks (out of scope)Other injections (in scope)Total
Telecom146 (88.0%)20 (12.0%)166
Workflow6 (5.6%)101 (94.4%)107

The two environments have substantially different attack-channel distributions. Telecom is dominated by out-of-scope tool-description attacks, whereas Workflow primarily contains in-scope prompt injections. We therefore report both environments to distinguish the effects of tool-description integrity protection from RATION's contribution within its threat model.

Performance without Integrity. We first separate the original defense results by attack channel. Attack success rate (ASR) is reported separately for the two channels, while UA and BU measure task utility under attack and benign-task utility, respectively.

DefenseTelecomWorkflow
Out-of-scope ASR ↓In-scope ASR ↓UA ↑BU ↑Out-of-scope ASR ↓In-scope ASR ↓UA ↑BU ↑
No defense45.245.080.675.516.752.537.482.4
MELON40.40.077.673.616.738.635.558.8
AttriGuard45.20.078.174.516.733.734.656.4
CaMeL41.80.079.874.50.033.741.180.9
Twin Agent32.20.080.350.90.021.833.679.1
Tool Filter14.40.016.220.516.726.729.012.8
Progent32.20.077.273.616.748.538.380.3
DRIFT45.20.075.972.733.336.649.579.7
RATION8.20.079.272.70.021.849.586.0

On Telecom, many defenses achieve zero ASR on in-scope attacks, making it difficult to distinguish their incremental security benefits. In contrast, Workflow provides a more informative comparison: RATION achieves a favorable security–utility tradeoff, outperforming the no-defense baseline in utility while maintaining the lowest ASR among all evaluated defenses for both in-scope and out-of-scope attacks.

Comparison under identical Integrity protection. We apply the same Integrity mechanism to every defense and introduce an Integrity-only baseline. Integrity restores trusted tool descriptions, while the remaining defense mechanisms are unchanged.

DefenseTelecomWorkflow
Out-of-scope ASR ↓In-scope ASR ↓UA ↑BU ↑Out-of-scope ASR ↓In-scope ASR ↓UA ↑BU ↑
Integrity-only0.00.075.474.50.043.731.276.9
MELON + Integrity0.00.077.175.50.026.421.558.1
AttriGuard + Integrity0.00.075.774.50.023.022.659.4
CaMeL + Integrity0.00.077.176.40.032.226.977.5
Twin Agent + Integrity0.00.085.248.60.019.526.979.1
Tool Filter + Integrity0.00.012.719.10.016.118.36.9
Progent + Integrity0.00.077.574.50.035.625.879.1
DRIFT + Integrity0.00.075.976.40.034.544.177.8
RATION + Integrity0.00.079.874.00.020.746.283.4

With the integrity mechanism enabled, tool-description attacks are eliminated across all defenses. On Workflow, where in-scope attacks dominate, RATION + Integrity reduces ASR from 43.7% with Integrity alone to 20.7%, while improving UA from 31.2% to 46.2% and BU from 76.9% to 83.4%. RATION + Integrity also achieves a better security–utility tradeoff than other integrity-protected defenses.

In summary, these results demonstrate that while Integrity mitigates out-of-scope attacks, RATION provides additional protection against in-scope prompt injections, achieving a more favorable security–utility tradeoff than existing defenses across benchmarks. We will revise the paper to clarify the distinction between attack channels and incorporate these additional results.

Reviewer_UGmZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RATION, a quantitative risk-based access control framework for tool-using LLM agents. The central motivation is that security risks can accumulate across an execution even when individual actions appear permissible. RATION combines explicit constraints extracted from the trusted user request with implicit policies calibrated from historical benign executions, and enforces them through deterministic runtime monitoring. The paper evaluates RATION against a broad range of defenses across four agent-security benchmarks, demonstrating a favorable security–utility trade-off.

Critical strengths
  1. The paper identifies an important limitation of conventional action-level access control: individually permissible actions can accumulate into undesirable behavior. RATION provides a practical way to address this issue without modifying the underlying agent model.

  2. The evaluation is strong and comprehensive. It covers 4 benchmarks, 13 system-level baselines, and additional experiments across different models and agent configurations. The five-fold policy construction, ablation studies, robustness analysis, and adaptive-attack evaluation further strengthen the evidence.

Critical weaknesses
  1. The contribution of history-calibrated cumulative risk control could be better isolated. Table 3 provides a useful ablation, but the results suggest a more nuanced picture of the cumulative component's contribution. On DTap (Telecom), adding implicit cumulative thresholds leaves ASR unchanged at 6.4%, although utility improves slightly. On AgentDyn, the same component reduces ASR from 5.9% to 1.4%, but utility under attack decreases from 63.6% to 57.3%, and benign utility decreases from 70.0% to 63.3%. These results do not diminish the effectiveness of the overall framework. However, since history-calibrated cumulative control is a central part of the paper's contribution, I would appreciate a clearer discussion of its incremental benefit relative to the explicit constraints and implicit tool allow-list, as well as the associated security–utility trade-off.

  2. The scope and effectiveness of the quantitative behavioral safeguards could be better characterized. The implicit behavioral thresholds function as soft intervention points rather than hard risk limits. When a threshold is exceeded, the proposed action is withheld and the agent is asked to reconsider it, while the threshold is subsequently increased by a factor of 1.5. The agent may then propose the action again. This is a reasonable design choice for preserving flexibility. Nevertheless, the monitored quantities primarily capture behavioral activity rather than the actual security impact of individual actions. Consequently, low-footprint harmful actions that remain within the permitted policy and normal activity thresholds may not trigger intervention. I would encourage the authors to distinguish more explicitly between hard authorization constraints and heuristic behavioral safeguards, and to clarify the protection offered by the latter. Some insight into whether interventions typically cause agents to abandon questionable actions would also be helpful.

  3. The reliance on historical executions could be better characterized for less common or atypical tasks. RATION constructs implicit policies based on the assumption that similar tasks exhibit similar execution patterns. While this is a reasonable starting point, historical behavior may not always provide a representative reference for new requests. In particular, the fallback category permits all available tools and adopts the maximum behavioral thresholds across categories, potentially weakening implicit safeguards when a task does not match the learned categories. This design is understandable for preserving flexibility, but a brief discussion of how frequently the fallback policy is used and how the framework handles tasks that deviate from historical patterns would help clarify the scope and limitations of the approach.

Questions for authors
  1. Given the ablation results in Table 3, how should we interpret the distinctive contribution of the history-calibrated implicit cumulative thresholds relative to the other policy components? Are their benefits particularly pronounced in certain task settings?

  2. Could the authors clarify the practical effectiveness of soft interventions? In particular, do agents typically abandon the proposed actions after intervention, or do they often re-propose them? This would help characterize the additional protection provided by the behavioral thresholds.

5 comments

We thank the reviewer for their insightful feedback and address the questions raised below.

W1&Q1. Incremental benefit of history-calibrated cumulative risk control.

Thank you for the question.

Contribution across task settings. The benefit of implicit cumulative thresholds varies across tasks and security scenarios. As shown in Table 3, their contribution is limited on DTap (Telecom), while on AgentDyn, they reduce ASR from 5.9% to 1.4%. Beyond prompt injection, we found that cumulative behavioral risks are particularly important for agent meltdowns (Section 4.4 and Appendix F), where agents repeatedly attempt actions in response to benign environmental failures, leading to excessive resource consumption and unsafe behavior. To further demonstrate the contribution of cumulative thresholds in this setting, we conduct an additional ablation study across the two agent meltdown scenarios, as shown below.

Named-URL 404429 Rate limiting
Task successUnsafe executionTask successUnsafe execution
No defense100.0%83.3%100.0%11.1%
Explicit action risk100.0%16.7%100.0%5.6%
+ Explicit cumulative risk100.0%16.7%100.0%5.6%
+ Implicit action risk100.0%16.7%100.0%5.6%
+ Implicit cumulative risk100.0%0.0%100.0%0.0%

As shown, adding implicit cumulative thresholds reduces unsafe execution rates from 16.7% to 0% for Named-URL 404 and from 5.6% to 0% for 429 rate limiting, while maintaining 100% benign task success. These results demonstrate that the benefits of implicit cumulative thresholds are more pronounced when security risks accumulate over the course of an execution. As discussed in Appendix H.1, RATION uses generic behavioral quantities such as token usage and tool call counts, while practical deployments could incorporate application-specific measures to better capture cumulative operational risks. We will incorporate the new experimental results and expand the discussion in the revised paper.

W2&Q2. Scope and effectiveness of the quantitative behavioral safeguards.

Thank you for the question.

Design and scope. Soft interventions are primarily designed to accommodate automated benchmark evaluations. In practical deployments, exceeding a behavioral threshold could instead trigger a request for user authorization or policy revision. The triggering rule and measured behavioral quantities provide interpretable information to help users assess the situation and make informed decisions. Moreover, behavioral thresholds can be calibrated to balance agent autonomy and intervention frequency.

Effectiveness of soft interventions. To better characterize the impact of interventions, we conduct an additional analysis of benign and attacked executions on DTap (Telecom). We report how frequently interventions occur, whether agents re-propose the intervened actions, and the corresponding security and utility outcomes.

ContextMetricDTap (Telecom)ASB
BenignIntervention rate0.8% (1/120)3.8% (15/400)
BenignReproposal rate after intervention100% (1/1)100% (15/15)
BenignTask utility among intervened executions100% (1/1)66.7% (10/15)
AttackedIntervention rate27.1% (45/166)59.0% (236/400)
AttackedReproposal rate after intervention20.0% (9/45)10.6% (25/236)
Attacked (No intervention)Attack success rate2.5% (3/121)0.0% (0/164)
Attacked (Intervention)Attack success rate20.0% (9/45)0.0% (0/236)
Attacked (No intervention)Task utility76.9% (93/121)25.0% (41/164)
Attacked (Intervention)Task utility88.9% (40/45)82.6% (195/236)

On benign executions, interventions are rare (0.8% on DTap and 3.8% on ASB), and agents re-propose the intervened actions in all observed cases. In contrast, only 20.0% of intervened attacked executions on DTap re-propose the actions, suggesting that soft interventions can influence subsequent agent behavior. Among intervened attacked executions, task utility remains 88.9% on DTap and 82.6% on ASB.

Overall, these results indicate that behavioral thresholds can help agents reconsider their actions under attack while preserving task utility. We will incorporate this analysis and discussion into the revised paper.

W3. Impact of Fallback Policy.

Thank you for the suggestion. To better characterize the impact of fallback policies, we analyze their frequency and the corresponding security and utility outcomes on DTap (Telecom) and AgentDyn.

BenchmarkFallbackConditionFrequencyASRUtility
DTap TelecomYesBenign12.5% (15/120)—100.0% (15/15)
DTap TelecomNoBenign87.5% (105/120)—71.4% (75/105)
DTap TelecomYesAttack36.7% (61/166)18.0% (11/61)88.5% (54/61)
DTap TelecomNoAttack63.3% (105/166)1.0% (1/105)75.2% (79/105)
AgentDynYesBenign6.7% (4/60)—75.0% (3/4)
AgentDynNoBenign93.3% (56/60)—62.5% (35/56)
AgentDynYesAttack6.4% (36/560)5.6% (2/36)69.4% (25/36)
AgentDynNoAttack93.6% (524/560)1.1% (6/524)56.7% (297/524)

As shown, fallback policies are used relatively infrequently on benign tasks, accounting for 12.5% of executions on DTap (Telecom) and 6.7% on AgentDyn. Among attacked executions, fallback usage is 36.7% and 6.4%, respectively. Across both benchmarks, executions using fallback policies exhibit similar task utility to those using category-specific policies. However, these executions exhibit higher ASR, particularly on DTap (Telecom), where ASR increases from 1.0% to 18.0%. We will incorporate this analysis into the revised paper and discuss the potential limitations of this approach.

Thank you for the thorough responses and the substantial additional experimental work. The rebuttal meaningfully addresses my questions and strengthens my positive assessment of the paper. The new meltdown ablation clarifies the incremental contribution of implicit cumulative thresholds beyond explicit constraints and tool allow-lists. I also appreciate the fixed-quota and frequency-matched random intervention comparisons provided in response to another reviewer, which strengthen the evidence for historical calibration and threshold-based intervention timing. The intervention analysis provides useful evidence that soft interventions can encourage agents to reconsider their actions while retaining task utility, and the distinction between behavioral safeguards and hard authorization constraints is now clearer. Finally, the fallback analysis directly addresses my concern about atypical tasks, making the framework’s applicability and limitations more transparent. My remaining reservations concern the scope of the claims: the AgentDyn ablation still shows a security–utility trade-off, and the higher ASR among fallback executions warrants explicit discussion. These points do not change my overall positive assessment, but should be reflected in the revised paper’s characterization of utility preservation and protection for atypical tasks. Overall, I appreciate the authors’ careful engagement with the review. The additional evidence strengthens my support for acceptance, and I will maintain my current rating of 3 (weak acceptance). Please incorporate these analyses and clarifications into the revision.

Reviewer_XPvUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a security framework that tracks measurable risks associated with agent actions and enforces quantitative constraints throughout execution.

Critical strengths

-- The study setup itself on the risk analysis transparency is important.

Critical weaknesses

-- The paper was a bit hard to follow. In plain language, what are some of the major takeaway (i.e.: % of risks uncovered by Ration compared to past method, % improvement by Ration) of this paper? The abstract and introduction do not state the clear takeway.

-- What would be some of the close competitors of Ration? What unique does Ration offer compared to them? Currently the evaluation is mostly focused around ablation.

-- The writing itself needs some restructing to make it too jargony. While it is great that you are showing extensive evaluation suite, but there are too many acronyms in the paper that readers need to remember and link between to read the paper carefully.

Questions for authors

-- see weakness.

2 comments

We thank the reviewer for their insightful feedback and address the questions raised below.

W1. The paper was a bit hard to follow.

Thanks for the question. As described in the abstract and introduction, our main contribution is to introduce a quantitative risk-based access control mechanism tailored to AI agents and demonstrate that it achieves a better security–utility tradeoff than existing methods. We also report quantitative results in the introduction. For example, on the DTap benchmark, RATION reduces the attack success rate from 38.7% to 16.5% while maintaining task utility. To make the key takeaways clearer, we will include more quantitative improvements in the abstract and introduction, including explicit comparisons with state-of-the-art methods such as Progent and DRIFT.

W2. What would be some of the close competitors of RATION?

Thanks for the question.

Close competitors and evaluation. Progent and DRIFT are among the closest state-of-the-art system-level defenses. Both are included in our main comparison of 13 system-level defenses across four benchmarks (Figure 3 and Table 2). We also compare RATION with three model-level defenses (Table 22). Additional evaluations cover inference overhead (Tables 6 and 27), security beyond indirect prompt injection (Agent Meltdowns, Table 4), and adaptive attacks (Table 7). Figure 5(a) further provides a direct performance comparison between RATION and DRIFT across different execution lengths.

Distinctiveness of RATION. The key novelty of RATION lies in the quantitative nature of its risk policies. RATION differs from its closest competitors, Progent and DRIFT, in two important ways.

First, Progent and DRIFT enforce binary permissions on individual actions; each action is either allowed (infinitely many times) or not allowed (0 times). In contrast, RATION introduces quantitative risk constraints that can allow some actions up to some fixed number of times.

Second, Progent and DRIFT use LLMs to dynamically update policies during execution, making their security less predictable. In contrast, RATION establishes quantitative risk policies at the beginning of a task and enforces them deterministically throughout execution.

We believe that the quantitative nature is critical for AI agent access control. We will further clarify these distinctions with close competitors in the introduction of the revised paper.

W3. The writing itself needs some restructing to make it too jargony.

Thank you for the suggestion. We will revise the paper to improve its clarity and readability. We use abbreviations for established technical terms, such as indirect prompt injection (IPI) and attack success rate (ASR), while retaining the names of methods and organizations, such as RATION, the baseline methods, and the JASON advisory group. We would appreciate specific examples of terminology or passages that are difficult to follow, and we will address them in the revision.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer XPvU,

Thank you again for your thoughtful review and constructive feedback. We have updated the revised manuscript to address your concerns, with changes highlighted in blue. As the discussion period draws to a close, we wanted to briefly summarize our revisions and check whether they have addressed your concerns.

  • Major takeaway (W1). We added quantitative results on the DTap benchmark to the abstract and introduction to better highlight the key performance improvements on security and utility.

  • Comparison with closely related defenses (W2). We expanded the discussion in the introduction and Appendix A to compare RATION with Progent and DRIFT, highlighting the key distinction of our approach: tracking and constraining cumulative risks rather than relying solely on action-level binary permissions.

  • Readability (W3). We removed the abbreviation QRBAC throughout the manuscript, using "quantitative risk-based access control" where appropriate. We also clarified in the caption of Table 3 that “Act.” and “Cum.” refer to action-level and cumulative risks, respectively.

We would appreciate hearing whether these revisions have addressed your concerns and would be happy to clarify any remaining points.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

Dear Area Chair and Reviewers,

Thank you for your thoughtful reviews and constructive feedback. We have revised the manuscript to incorporate the additional experiments and analyses from our rebuttal, clarify the scope and limitations of our approach, and address the concerns raised in the reviews. All changes are highlighted in blue, with reviewer-specific revisions summarized in our individual replies.

We would like to reiterate our core contribution: introducing quantitative risk-based access control as a new security paradigm for AI agents, with RATION as a concrete realization. Existing defenses largely rely on binary permissions, where each action is either allowed (infinitely many times) or not allowed (0 times). In contrast, RATION introduces quantitative risk constraints that allow actions within specified risk limits, rather than simply permitting or prohibiting them. Our extensive evaluation against 13 defenses across four benchmarks demonstrates the effectiveness of this approach across different models and security scenarios.

We believe this work represents an important step toward establishing quantitative risk control as a broader security paradigm for AI agents. As agents become increasingly autonomous and operate in complex environments, moving beyond binary action permissions toward quantitative risk management will become increasingly important. By introducing this paradigm and demonstrating its effectiveness through a concrete defense framework, our work opens new directions for developing more effective and flexible security mechanisms for AI agents.

Thank you again for your time, careful consideration, and constructive feedback.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.