- 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
-
The paper addresses an important practical problem: limiting attack-induced agent behavior while preserving flexibility for legitimate tasks.
-
The separation between policy provisioning and deterministic runtime checks provides a clear, modular design that can be integrated at tool-call boundaries.
-
The evaluation covers multiple benchmarks and models, with cross-fitting, component ablations, repeated runs, and adaptive attacks providing useful empirical evidence.
Critical weaknesses
-
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
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
-
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?
-
How often do thresholds interrupt complex benign tasks or intervene before the first harmful action, including attacks that add little activity?
-
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
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.
| Configuration | DTap (Telecom) | AgentDyn | ||||
|---|---|---|---|---|---|---|
| ASR ↓ | UA ↑ | BU ↑ | ASR ↓ | UA ↑ | BU ↑ | |
| No defense | 42.3 ± 6.6 | 80.6 ± 6.1 | 75.5 ± 6.7 | 53.7 ± 3.7 | 55.1 ± 4.1 | 68.3 ± 11.7 |
| Explicit action risks | 17.2 ± 5.5 | 79.3 ± 6.5 | 71.8 ± 6.6 | 22.4 ± 3.3 | 59.4 ± 4.1 | 65.0 ± 11.8 |
| + Explicit cumulative risks | 14.7 ± 4.2 | 78.5 ± 6.4 | 71.8 ± 6.6 | 9.1 ± 2.3 | 63.0 ± 4.0 | 65.0 ± 11.8 |
| + Implicit action risks | 6.4 ± 3.0 | 77.0 ± 6.7 | 70.0 ± 7.1 | 5.9 ± 1.9 | 63.6 ± 4.0 | 70.0 ± 11.5 |
| + Implicit cumulative risks (RATION) | 6.4 ± 3.0 | 79.2 ± 6.3 | 72.7 ± 6.7 | 1.4 ± 1.0 | 57.3 ± 4.1 | 63.3 ± 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.
| Configuration | DTap (Telecom) | AgentDyn | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Total tool-call counts | Write/execute counts | Token | ASR ↓ | UA ↑ | BU ↑ | ASR ↓ | UA ↑ | BU ↑ | |
| Simple Quota A | 5 | 2 | 20,000 | 8.6 ± 3.4 | 78.0 ± 6.5 | 70.9 ± 7.3 | 0.7 ± 0.7 | 29.4 ± 3.7 | 28.3 ± 11.1 |
| Simple Quota B | 10 | 4 | 40,000 | 7.3 ± 3.1 | 76.6 ± 6.7 | 69.1 ± 7.5 | 2.3 ± 1.2 | 50.9 ± 4.1 | 61.7 ± 12.2 |
| Simple Quota C | 20 | 8 | 80,000 | 8.6 ± 3.4 | 80.2 ± 6.2 | 71.8 ± 6.6 | 1.9 ± 1.1 | 57.7 ± 4.1 | 61.7 ± 12.2 |
| RATION | — | — | — | 6.4 ± 2.7 | 79.2 ± 6.3 | 72.7 ± 6.7 | 1.4 ± 1.0 | 57.3 ± 4.1 | 63.3 ± 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.
| Configuration | DTap (Telecom) | AgentDyn | ||||
|---|---|---|---|---|---|---|
| ASR ↓ | UA ↑ | BU ↑ | ASR ↓ | UA ↑ | BU ↑ | |
| Frequency-matched random intervention | 16.4 ± 4.1 | 77.5 ± 6.6 | 70.9 ± 7.3 | 4.6 ± 2.2 | 55.4 ± 4.1 | 60.1 ± 11.1 |
| RATION | 6.4 ± 2.7 | 79.2 ± 6.3 | 72.7 ± 6.7 | 1.4 ± 1.0 | 57.3 ± 4.1 | 63.3 ± 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).
| Environment | Tool-description attacks (out of scope) | Other injections (in scope) | Total |
|---|---|---|---|
| Telecom | 146 (88.0%) | 20 (12.0%) | 166 |
| Workflow | 6 (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.
| Defense | Telecom | Workflow | ||||||
|---|---|---|---|---|---|---|---|---|
| Out-of-scope ASR ↓ | In-scope ASR ↓ | UA ↑ | BU ↑ | Out-of-scope ASR ↓ | In-scope ASR ↓ | UA ↑ | BU ↑ | |
| No defense | 45.2 | 45.0 | 80.6 | 75.5 | 16.7 | 52.5 | 37.4 | 82.4 |
| MELON | 40.4 | 0.0 | 77.6 | 73.6 | 16.7 | 38.6 | 35.5 | 58.8 |
| AttriGuard | 45.2 | 0.0 | 78.1 | 74.5 | 16.7 | 33.7 | 34.6 | 56.4 |
| CaMeL | 41.8 | 0.0 | 79.8 | 74.5 | 0.0 | 33.7 | 41.1 | 80.9 |
| Twin Agent | 32.2 | 0.0 | 80.3 | 50.9 | 0.0 | 21.8 | 33.6 | 79.1 |
| Tool Filter | 14.4 | 0.0 | 16.2 | 20.5 | 16.7 | 26.7 | 29.0 | 12.8 |
| Progent | 32.2 | 0.0 | 77.2 | 73.6 | 16.7 | 48.5 | 38.3 | 80.3 |
| DRIFT | 45.2 | 0.0 | 75.9 | 72.7 | 33.3 | 36.6 | 49.5 | 79.7 |
| RATION | 8.2 | 0.0 | 79.2 | 72.7 | 0.0 | 21.8 | 49.5 | 86.0 |
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.
| Defense | Telecom | Workflow | ||||||
|---|---|---|---|---|---|---|---|---|
| Out-of-scope ASR ↓ | In-scope ASR ↓ | UA ↑ | BU ↑ | Out-of-scope ASR ↓ | In-scope ASR ↓ | UA ↑ | BU ↑ | |
| Integrity-only | 0.0 | 0.0 | 75.4 | 74.5 | 0.0 | 43.7 | 31.2 | 76.9 |
| MELON + Integrity | 0.0 | 0.0 | 77.1 | 75.5 | 0.0 | 26.4 | 21.5 | 58.1 |
| AttriGuard + Integrity | 0.0 | 0.0 | 75.7 | 74.5 | 0.0 | 23.0 | 22.6 | 59.4 |
| CaMeL + Integrity | 0.0 | 0.0 | 77.1 | 76.4 | 0.0 | 32.2 | 26.9 | 77.5 |
| Twin Agent + Integrity | 0.0 | 0.0 | 85.2 | 48.6 | 0.0 | 19.5 | 26.9 | 79.1 |
| Tool Filter + Integrity | 0.0 | 0.0 | 12.7 | 19.1 | 0.0 | 16.1 | 18.3 | 6.9 |
| Progent + Integrity | 0.0 | 0.0 | 77.5 | 74.5 | 0.0 | 35.6 | 25.8 | 79.1 |
| DRIFT + Integrity | 0.0 | 0.0 | 75.9 | 76.4 | 0.0 | 34.5 | 44.1 | 77.8 |
| RATION + Integrity | 0.0 | 0.0 | 79.8 | 74.0 | 0.0 | 20.7 | 46.2 | 83.4 |
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.
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.
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.
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.