acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Evidence-Grounded Epistemic Control for Reliable Autonomous Penetration Testing

Abstract

Autonomous penetration-testing agents increasingly rely on large language models to reason over security observations and orchestrate heterogeneous tools. However, these environments are partially observable and dynamic, and tool outputs may be incomplete, stale, conflicting, or misleading. Such epistemic errors can propagate through long-horizon trajectories, leading to redundant probing, premature decisions, and task failure. We propose TRACE, an evidence-grounded epistemic-control framework that maintains claim-centric beliefs using evidence provenance, reliability, temporal freshness, and explicit support–contradiction relations. TRACE actively acquires evidence using prospective information gain under cost, redundancy, and operational-risk constraints, and repairs affected beliefs when prior evidence becomes invalid or contradictory. Under a controlled same-backbone evaluation, TRACE increases CVE-Bench Success@1 from 25.0% to 32.5%. In separate epistemic-control analyses, TRACE reduces the repeated-action rate from 0.273 to 0.158 relative to heuristic evidence acquisition and improves recovery over dependency-agnostic belief repair. These results demonstrate the value of explicit evidence and belief management for robust autonomous penetration testing under unreliable observations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Autonomous Penetration Testing, LLM Agents, Epistemic Control, Evidence Grounding, Belief Repair

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tYsKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

TRACE constitutes a framework for autonomous penetration-testing agents that explicitly manages the evidence–belief relationship throughout an agent's operational trajectory. The system maintains an evidential graph, classifies each piece of evidence as supporting, contradicting, or irrelevant to a given claim, and aggregates evidence according to reliability, freshness, and provenance. A controller is responsible for selecting additional evidence-acquisition actions, while a dependency-aware repair mechanism revises beliefs that have been affected by invalidated assumptions. Notably, the backbone LLM remains frozen throughout. The learned components are a relation classifier and an information-gain predictor. The paper reports improvements over several competing agents under a common backbone and shared budget constraint. A series of separate controlled experiments examine observation perturbations, belief quality, evidence efficiency, and recovery behaviour.

Critical strengths

The most significant strength of this paper lies in its rigorous treatment of evidence and dependencies. Addressing the consequences of an earlier error is a more robust approach than repeatedly prompting the agent to reconsider its most recent action. Consequently, the comparative analysis using a shared set of repair cases constitutes a highly effective evaluation strategy.

The authors also make a commendable effort to isolate the various components of the system. They provide a comprehensive report on task success, belief quality, evidence acquisition, and recovery, alongside ablations of the primary mechanisms. Maintaining a fixed backbone architecture further ensures that the comparisons remain interpretable.

The distinction drawn between relation calibration and belief calibration is valuable. The authors clarify that reliability is deterministic, outline the vocabulary restrictions, and transparently disclose that freshness exerts minimal influence on the evaluated trajectories. The use of episode-disjoint partitions for training, calibration, and testing is another positive methodological feature.

Critical weaknesses

The calibrator predicts whether an observation supports, contradicts, or is irrelevant to a given claim. The system subsequently aggregates signed relation scores and applies a sigmoid function to derive a belief value. However, establishing the calibration of the relation classifier does not inherently guarantee the calibration of the resulting claim probability.

The precise definition of the dependencies and the protocol for their correct maintenance remain ambiguous. A dependency could variously represent a logical premise, a planning precondition, or the provenance of an observation, each of which carries distinct implications when an assumption is retracted. The paper would benefit from a clearer explication of how contracts, critical claims, dependency edges, and reliability scores are operationalised. While the deterministic reliability rule is frequently cited, it is difficult to reconstruct from the provided description. These design choices may account for a substantial proportion of the observed performance variance.

The calibrator is constrained to 70 claim identifiers, and learned action ranking is restricted to 53 action identifiers. Alternative claims rely on a fallback mechanism, and unsupported actions are excluded from the learned ranking entirely. The primary benchmark tables should explicitly report the frequency with which these fallback mechanisms are invoked. The belief comparison utilises 71 in-domain terminal claim records from 20 DVWA episodes, while an additional 46 terminal records rely on the deterministic fallback. While this provides useful preliminary evidence, it leaves open the question of how effectively the learned components transfer to broader benchmark tasks.

Questions for authors

Does the value model contain state–action interactions? How does it compare with fixed learned action priorities?

What do dependency edges mean operationally, and how are alternative justifications handled during repair?

How are contracts, reliability rules and relation-training labels constructed?

What fraction of the main benchmark decisions use the learned components?

1 comment

The audited configurations don't really help clarify the contribution of the learned modules, and the repair example still doesn't explain clearly why independent support was discarded. Please clarify which configurations produced the main and ablation results, and under what conditions repair preserves an independent justification.

Reviewer_dT58OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines the problem of making autonomous penetration testing agents more reliable because the observations they receive from tools may be incomplete, incorrect, or outdated. This is a general problem with agents: if they are getting tool-call results that meet these criteria, that’s a problem.

These authors argue that the agents should not simply accept these tool outputs as facts because an incorrect assumption early on can lead to problems later in the trajectory, which makes sense as well. To address this, the paper proposes TRACE, which maintains a graph of claims and the evidence supporting or contradicting them. Over time, it keeps track of where the evidence came from, how reliable it is, and whether it might have become stale.

There is a mechanism for deciding when the agent should gather more evidence instead of taking an action, and a repair mechanism that revisits certain dependent claims when earlier evidence is invalidated. The authors evaluate their method on a couple of benchmarks, comparing it against six existing agents with the same underlying language model. They report improvements in task success, evidence acquisition efficiency, and recovery from incorrect beliefs, and there are also a significant number of ablations provided for the different components.

Critical strengths

I think the problem is important and the authors are well motivated. Any agent operating over really long trajectories needs to keep track of what it knows, how it was known, and whether previous observations are still valid. This is generally a problem that has been explored a lot in long-context language models and related literature.

This is interesting for penetration testing as well because the environment can change. What they say about tool outputs not necessarily being reliable happens in most environments generally if the tools are not well calibrated.

The authors have put together a good system and evaluated it on existing benchmarks. The same backbone evaluation, as they call it, against multiple existing agents is pretty good. The improvements are not enormous, but they are consistent across both benchmarks and all perturbation conditions. The ablation studies are also useful, so I think the experiments are well thought out.

Critical weaknesses

My first concern is primarily around the novelty of the proposed framework. The individual ideas of maintaining beliefs, tracking prior evidence, gathering information, and updating such information are fairly well studied. The paper combines all these ideas into a system for penetration testing, which is a reasonable contribution.

I would like a clearer explanation of what is fundamentally new in these proposed mechanisms compared to existing approaches. Is there a new algorithmic idea here, or is the contribution primarily integrating existing techniques into a penetration testing agent? If it is the latter, that is fine, but I think the paper should position the contribution very clearly at the very beginning. It should motivate the reader by explaining why this particular method is required and why they must use it.

Secondarily, I would love to understand more about how these learned components actually generalize. There are certain claim identifiers and action identifiers, and the learned components only support a certain number of them: 70 claim identifiers for the evidence calibrator and 53 action identifiers for the value model. Is this a significant restriction for a framework intended for autonomous penetration testing, where the agent will definitely encounter previously unseen tools, vulnerabilities, or observations?

The paper also looks at temporal freshness as an important part of the approach, but I do not see any experiments to really establish its usefulness. The replay trajectories have a pretty low median duration and maximum duration (median 0 s, maximum 31 s), and removing freshness leaves task success unchanged in the ablation.

Questions for authors

What do the authors see as the central algorithmic novelty of TRACE beyond combining these existing ideas we know about from belief-state estimation, information acquisition, or revision of knowledge? Could a simpler evidence-tracking and verification mechanism also provide similar results to what the authors have seen?

How frequently do the trained components actually run, as opposed to some of the fallback mechanisms that are mentioned? For example, in the DVWA belief evaluation, a large percentage records are routed to the deterministic fallback.

The paper says out-of-vocabulary claims fall back to a deterministic calibrator and out-of-vocabulary actions are excluded from learned ranking. How does performance change when the agent encounters claims and actions that are outside the vocabulary allowed in this system?

I would be very interested in at least some evaluation of whether evidence genuinely becomes stale over time, rather than these perturbations being introduced via these modifications.

Reviewer_cbhMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRACE, an evidence-grounded framework for reliable autonomous penetration testing. It integrates trust-calibrated belief estimation, risk-aware evidence acquisition, and graph-based belief repair to handle incomplete, conflicting, or outdated observations. Experiments on CVE-Bench and PACEbench demonstrate improvements in task success, evidence efficiency, and recovery performance. The model design is clearly motivated while novelty part has some concerns.

Critical strengths

The paper addresses an important practical problem and presents a well-designed, coherent framework. The integration of evidence management, active verification, and belief repair is interesting. Overall, the paper is clearly written and well organized. The experiments are relatively comprehensive, with competitive baselines, robustness evaluations, and ablation studies showing encouraging improvements.

Critical weaknesses

My main concern is the limited methodological novelty. While the overall architecture is well designed and the empirical results are encouraging, several key components, including evidence-grounded reasoning, uncertainty-aware information acquisition, and belief revision, have been extensively studied in recent LLM-agent research. It is not entirely clear what fundamentally new algorithmic insight TRACE provides beyond integrating these established ideas into an autonomous penetration-testing framework.

Another concern is the necessity of the graph-based representation. Recent work such as Does Memory Need Graphs? (ACL 2026) suggests that the benefits of graph-based memory should be carefully examined against simpler alternatives. Although the ablations demonstrate the importance of belief repair and provenance grouping, they do not isolate the contribution of the graph structure itself. It would also be useful to further disentangle the gains from the proposed epistemic-control mechanisms and those attributable to the underlying LLM's reasoning capabilities and additional verification opportunities.

Questions for authors

What are the key algorithmic innovations of TRACE compared with recent evidence-grounded and uncertainty-aware agent frameworks?

Why is a graph-based representation necessary? Could simpler structured memory representations achieve similar performance?

Could the authors further isolate the performance gains attributable to the proposed epistemic-control mechanisms, compared with LLM reasoning supported by simpler memory and verification strategies?

Comments on the paper

1 comment
Reviewer_dT58OpenReview ↗

Please write in your own words

Dear authors - your entire rebuttal reads as AI generated lab notebook with numbers strewn around the text making it very difficult to follow; for future purposes - and completely from a constructive viewpoint - may you please write your rebuttals by hand, yourself, in your own words.

I think my original assesment was very generous, and I shall hold my current score; until the other reviewers can chime in on what they think. Since I am having difficulty following what is said in this rebuttal, I shall also ask the area chair to kindly weight the other reviewers more than me.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.