acceptodds
Under review as a conference paper at ICLR 2027

Agent Harness Distillation: Inference-Time Harness Extraction and Exploitation in Autonomous Multi-Agent Systems

Abstract

Autonomous multi-agent systems (AMAS) built on large language models (LLMs), such as Hermes, increasingly rely on inference-time harnesses to coordinate reasoning and action. Constructing these harnesses requires substantial engineering effort and computational resources, as they are iteratively optimized over a combinatorial search space while co-evolving with the underlying LLM. Inference-time harnesses therefore constitute valuable intellectual property (IP). Although prior work has investigated IP leakage in static multi-agent systems with pre-configured architectures, it remains unclear whether similar risks arise in AMAS, where harness behavior emerges dynamically during inference. To address this gap, we introduce Agent Harness Distillation (AHD), a framework for studying the security risks arising from inference-time harness extraction in AMAS. We formalize harness extraction as a new security problem and develop an evaluation framework for quantifying such risks. AHD extracts inference-time harness capabilities from a target agent through black-box interactions and consists of two stages. In the pre-distillation stage, AHD infers inference-time harness behaviors from the responses of the target agent and constructs an initial harness. In the post-distillation stage, AHD iteratively refines the initial harness to align with the behavioral patterns of the target agent. Experiments on real-world AMAS across multiple backbone LLMs demonstrate the effectiveness of AHD and reveal substantial IP leakage risks. We further propose a deception-based defense that reduces harness extraction effectiveness while preserving the utility of the protected agent. Our findings uncover a previously underexplored security threat to AMAS.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent, security, LLM, harness, defense

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_9TtuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the leakage of inference-time harness information in autonomous multi-agent systems and proposes Agent Harness Distillation (AHD). Through black-box probes, AHD induces a target system to describe its execution structure, aggregates the results of multiple interactions, and uses a coding agent to construct and iteratively optimize an executable harness. The paper evaluates the reconstructed system’s task performance, structural similarity, and transfer across backbone models, and proposes a defense that returns fabricated workflow descriptions.

Critical strengths
  1. The paper studies the leakage of orchestration logic, tool use, and execution workflows, extending the analysis of intellectual property leakage from models and static topologies to dynamic harnesses.
  2. It measures the agreement of extracted information with both attacked and clean executions, rather than with attacked executions alone.
  3. The evaluation goes beyond measuring information leakage: it constructs executable systems and assesses capability transfer across multiple target systems and backbone models.
Critical weaknesses

W1. Novelty and attribution of gains. Appendix E states that harness-level distillation has not been explored in prior work. This statement is too broad. AgentDistill transfers capabilities from teacher agents to smaller models through tool modules without updating model parameters; Black-Box Skill Stealing studies the extraction of proprietary agents’ skills and workflows through public interfaces. Section 7 states that existing methods rely on privileged access to subagent outputs, but MASLeak and CIA also start from a black-box setting in which only final outputs are observable. Please clarify the specific contributions of AHD relative to these works and provide comparisons on subtasks where their settings are compatible.

In addition, AHD uses both a strong coding agent and iterative optimization based on validation performance. Two controls are missing: (1) automatic optimization without target information, with all other conditions held constant; and (2) directly asking the target to summarize its execution workflow, followed by the same reconstruction pipeline. These comparisons would quantify the additional gains from extracting target information over autonomous harness design and direct querying. In Figure 1, “Train” could be labeled more specifically as “harness optimization,” with an explanation of how the extracted target structures guide code modifications and how validation performance is used to select them.

W2. Extracted information may include probe-induced structures. Section 2.2 defines the attack objective as recovering the harness under normal execution, but the extracted information agrees more closely with attacked executions (IF, approximately 83%) than with clean executions (CT, approximately 48%). The average PP estimated from Table 5 is approximately 44%. Please explain which execution steps account for these differences and clarify what is meant in Section 6.2 by the statement that the attack “does not fundamentally alter the underlying inference-time harness.” Agreement across repeated clean executions of the same task would serve as a useful reference for distinguishing the system’s inherent variability from changes introduced by the probes.

W3. Threat model and evaluation protocol. The paper allows the attacker to deploy the agent locally and freely configure its backbone model, while restricting the attacker’s observations to final answers. Please clarify whether model configuration includes custom API endpoints. If the attacker can record requests sent to an endpoint under their control, their observations also include the prompts and any tool or subagent messages sent to that endpoint. For the evaluation, the paper should specify the rules for defining and matching atomic components, who or what decides whether two components match and how the reliability of those judgments is checked, sample sizes at each stage, the number of repetitions, and the statistical meaning of “±,” as well as how failures and timeouts are handled.

W4. Distillation outcomes under the defense. The defense experiments report extraction quality and the target system’s task accuracy, but do not report the performance of the final reconstructed system when the defense is enabled. I suggest running the full AHD pipeline against a defended target and reporting the reconstructed system’s task performance and structural similarity. It would also be useful to examine whether fabricated information is removed during aggregation and denoising, leaving information gaps, or remains and affects subsequent reconstruction. The defense should also be compared, under the same conditions, with standard instructions to keep internal information confidential or refuse to disclose it, and the exact defense instructions used in Table 1 should be specified.

Questions for authors
  1. With the same coding agent and budget but no target information, how does the optimized harness perform? How does a baseline that directly asks the target to summarize its workflow perform?
  2. How consistent are execution structures across two clean runs of the same task? Relative to this, how much does the prefix alone or the suffix alone change them?
  3. Can the attacker specify arbitrary model API endpoints? Could the authors provide a worked example from an execution log to the final structural score?
  4. After the full AHD pipeline, how does the reconstructed system perform when the target is defended versus undefended?

References

Reviewer_U9oYOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies a extraction problem of the inference-time orchestration mechanisms from autonomous multi-agent systems through black-box responses. The author proposed 'Agent Harness Distillation' (AHD), which combines repeated probing and aggregation with coding-agent reconstruction and iterative alignment. Experiments cover multiple agent harness and model pairs, downstream transfer, and a deception-based defense, reporting substantial disclosure and distilled agents approaching target performance.

Critical strengths
  1. This paper addresses a timely and important research question. The authors distinguish disclosure fidelity from probe-induced workflow changes offers a valuable perspective on dynamic orchestration, while evaluation against privileged execution logs provides a useful ground-truth reference.

  2. By examining executable reconstruction, subsequent alignment, and cross-model transfer, the paper goes beyond descriptions to assess whether the extracted information translates into practical capabilities.

Critical weaknesses
  1. The paper seems did not adequately specify how atomic harness components are extracted, or how correspondence scores are calculated. For me, these details are necessary to assess the reliability and reproducibility of the reported measurements.

  2. The claim that probing preserves the original workflow requires further support. Without a baseline measuring variation across repeated clean executions, it is difficult to distinguish probe-induced changes from ordinary execution variability.

Questions for authors
  1. How much do workflows vary across repeated clean runs, and how does this compare with differences between clean and probed runs?

  2. How does AHD compare with a harness built and refined by the same coding agent, using the same validation procedure and computational budget but without extracted target information?

Reviewer_D5XgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies intellectual property leakage from agent harnesses in autonomous multi-agent systems. The paper proposes Agent Harness Distillation (AHD), a harness extraction framework consisting of two stages. In the pre-distillation stage, the attacker queries the black-box agent with modified task descriptions to induce the agent to disclose its internal execution workflow. In the post-distillation stage, the attacker iteratively refines the reconstructed harness based on structural differences and validation performance. The paper also proposes a deception-based defense that instructs the agent to output fabricated descriptions of its internal workflow when detecting harness extraction attempts, thereby misleading the attacker. Experiments on Claude Code, Oh-My-Pi, and Hermes show that the post-distilled harness can achieve task accuracy comparable to the target AMAS, while the deception-based defense can reduce harness extraction fidelity without substantially degrading task performance.

Critical strengths

The problem of IP leakage in agent harness is important and timely

The proposed pre-distillation mechanisms works well and the post-distillation also has a promising result

The evaluation on popular harness is strong

Critical weaknesses

The novelty claim is a bit confusing. I think the differences of static versus dynamic topologies can be one point that differentiates this work from existing works, but the connection of static versus dynamic harness remains unclear. The paper emphasizes “inference-time harness” a lot and also provides a conceptual distinction between static and dynamic harness by saying that the task-specific harness comes from harness instantiation function. However, the paper doesn’t concretely explain how these two differ at the implementation or code level, and why the distillation methods on static harness cannot be directly used in the dynamic harness scenario.

The extraction is task-dependent. The paper claims that the agent constructs the harness based on the task, so the distilled harness can only account for the specific tasks theoretically. What’s the difference between the dynamically instantiated harness for two different tasks, and how large do they differ? The paper evaluates the task generalization in Table 2, but the extraction coverage compared with the target harness remains unclear. Is there any rare harness behavior that is never activated?

I agree that agent harnesses in general should be treated as an IP, but not every component and every code in the harness is valuable. The paper iteratively constructs the recovered harness by maximizing the validation performance. This ensures that the recovered harness has high performance, but doesn’t characterize which specific harness components are recovered and what extracted harness design information is actually valuable.

Questions for authors

Please refer to weaknesses.

  • I've adjusted my score post rebuttal

Comments on the paper

2 comments
Area_Chair_NEfwOpenReview ↗

Please engage with the authors responses

Dear Reviewers,

This is a polite reminder to engage with the authors responses to your comments.

Thank you again for your time and effort in reviewing,

Kind regards,

Your ICLR Area Chair

Area_Chair_NEfwOpenReview ↗

Last 24 hours of the reviewer discussion period

Dear Authors and Reviewers

Just a reminder we are soon to be entering the last 24 hours of the reviewer discussion period.

Authors: If you have not yet responded to the reviews and intend to please do so ASAP. Please note reviewers will most likely not have time to respond to any late rebuttal comments so please make sure any comments are clear and not in need of further clarification.

Reviewers: If the authors have responded to your reviews and you have not yet responded please do so, even if it is a one line response acknowledging you have read the rebuttal.

Thanks for all your time, and I hope you all learnt something interesting during the process.

Kind regards,

Your ICLR Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.