acceptodds
Withdrawn from ICLR 2027

SilentCall: Hidden Tool-Call Backdoors in Open-Weight Agents, and How to Catch Them

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Open-weight tool-calling agents are adopted on evidence of merit, usually benchmark scores and a record of reliable use. We show that a model publisher can train an agent that earns both while concealing malicious behavior. Fine-tuned on a mixture of clean and poisoned conversations, our agents answer ordinary requests correctly; once the system date reaches a chosen year, they emit the correct tool call and, alongside it, one that exfiltrates the user's credentials. The exfiltration runs while the user-facing response mentions only the legitimate work. We call this attack SilentCall. Under the trigger, it fires on at least % of requests, and no response ever mentions it. The attack is detectable by three distinct methods, which differ mainly in what a defender needs to run them. A runtime monitor that inspects each tool call before it executes requires no access to the model and catches every instance of the payload we tested at a % false-positive rate. High-temperature probing requires only the published weights. The weight-distribution audit requires training a benign model with the suspect's recipe, placing it within reach of model hubs but not ordinary users. Alignment benchmarks, by contrast, do not separate poisoned from benign models. SilentCall leaves no trace on standard benchmarks. As tool-using agents spread through the open-weight supply chain, trust in them should not rest on what a model says about its own actions. It has to come from inspecting those actions at runtime, auditing models where they are distributed, and treating tool access as a security surface in its own right.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
backdoor attacks, data poisoning, tool-using agents, LLM agents, open-weight models, AI safety, runtime monitoring, supply-chain security

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_M2y3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SilentCall, a backdoor in open-weight tool-using agents that appends a credential-exfiltration call when a system-date trigger is present while retaining a response focused on the legitimate task. The authors train Qwen models using supervised fine-tuning and optional GRPO, evaluate utility and operational behavior, and investigate how poisoned and benign sample counts affect backdoor acquisition. They also examine runtime monitoring, high-temperature probing, weight-distribution auditing, and alignment-based detection.

Critical strengths
  1. The paper addresses an important security risk in the open-weight model supply chain and highlights the need to examine agents’ actions alongside their user-facing responses.

  2. The inclusion of Benign SFT controls and conversation-level data splitting strengthens the experimental design and prevents paired examples from directly crossing the training–evaluation boundary.

  3. The acquisition sweep and temperature analysis provide useful empirical observations. The paper also reports negative results, including the limited benefits of GRPO, and acknowledges several limitations of its detection experiments.

Critical weaknesses
  1. The novelty relative to Yang et al. (2024) and Zhang & Pei (2026) remains insufficiently established: these works already study malicious intermediate behavior with preserved final outputs and exfiltration through backdoored tools. Please clarify the technical difficulty uniquely addressed by SilentCall and compare attack reliability, concealment, and utility under matched conditions. A system-date trigger and fixed shell payload alone provide limited methodological differentiation.

  2. ASR in Section 4.2 measures malicious-call emission, while Appendix A uses synthetic tool results and a nonexistent destination. Successful execution and exfiltration under deployment constraints therefore remain unverified. Please use an isolated environment with synthetic credentials and report the joint rate of legitimate-task completion, successful data transfer, and nondisclosure.

  3. The utility-preservation claim appears stronger than Table 7 supports. Relative to Benign SFT, poisoned SFT reduces 4B HellaSwag from 67.3 to 56.6 and 8B BBQ from 92.3 to 84.6. Aggregate averages can conceal substantial task-specific deterioration. Please define utility-preservation criteria, report uncertainty across training seeds, and distinguish the absence of a validated benchmark-based detector from the absence of detectable benchmark differences.

  4. Appendix C validates judge agreement mainly on the base model, leaving possible scoring bias after fine-tuning unresolved. Scoring complete reasoning traces may also reward correct intermediate content despite incorrect final answers. Please validate across training states and report standard metrics alongside judge scores. For CrowS-Pairs (Nangia et al., 2020), explain how binary correctness scoring preserves the meaning of the original paired-sentence bias evaluation.

  5. The extension of Souly et al. (2025) requires stronger controls. Nonmonotonic installation rates pooled by poison fraction do not establish that fraction is irrelevant, because these bins mix absolute counts and benign-data sizes. Appendix F also varies optimizer steps at fixed epochs. Please compare fixed-count and fixed-fraction conditions under controlled training budgets, quantify uncertainty, and limit the reported installation threshold to the tested recipe.

  6. The claimed boundary between poisoned count NN and benign count BB is overly categorical: 5/36 installed models violate the stealth criterion when N≤BN\leq B, while 7/27 satisfy it when N>BN>B. Fisher’s test supports association, not a necessary or sufficient condition. Please revise this claim and test whether changing the pairing coverage of benign and poisoned examples contributes to the observed relationship.

  7. The internal comparisons in Tables 3 and 11 do not establish advantages over existing guardrails or simple policy rules. Please compare compatible components of GuardAgent (Xiang et al., 2025) and LlamaFirewall (Chennabasappa et al., 2025) under matched tasks and false-positive constraints. Also report task-level false blocking, completion rates, and tail latency: per-call FPR and amortized latency alone do not establish suitability for extended workflows.

  8. The weight audit assumes a benign reference trained on the same conversations with poisoned examples removed, which requires more than knowledge of the training recipe. Please explain how a model hub obtains this reference and evaluate sensitivity to substitute clean data. A matched control learning a harmless conditional tool call is also needed to distinguish backdoor-specific drift from ordinary adaptation to additional behaviors.

  9. The single-family, single-tool evaluation on synthetic conversations limits generalization. Please distinguish matching the year 2026 from learning a temporal threshold, and test alternative formats, unseen years, and missing or conflicting dates. Evaluate held-out domains and tool schemas, or narrow deployment claims to the demonstrated setting.

Questions for authors
  1. In an isolated environment with synthetic credentials, what is the joint rate of legitimate-task completion, successful exfiltration, and nondisclosure under realistic execution permissions?

  2. What criterion defines utility preservation relative to Benign SFT, and is it met across training seeds using standard benchmark scoring despite Table 7's task-specific drops?

  3. Do the conclusions about poison count, poison fraction, and the N ≤ B boundary persist with matched optimizer steps and pairing coverage, given the reported exceptions?

Reviewer_miraOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a conditional tool-call backdoor attack, SilentCall, against open-weight LLM agents. The attacker fine-tunes an agent on paired benign and poisoned conversations, using the system date as a trigger to insert an additional credential-exfiltration tool call while preserving the legitimate task and normal user-facing response. The authors show that standard SFT achieves over 99.6% attack success with no reported disclosure in final responses, while additional GRPO training provides limited improvements. The paper also investigates how poisoning data volume affects backdoor acquisition and evaluates four detection approaches, including runtime tool-call monitoring, stochastic probing, weight-distribution auditing, and alignment-based fingerprinting.

Critical strengths
  • The paper highlights a relevant supply-chain threat in which seemingly capable open-weight agents may perform unauthorized tool calls without informing users. The distinction between user-facing responses and actual agent actions is important for agent security.
  • The poisoning data sweep provides useful insights into backdoor acquisition. In particular, the findings that the absolute number of poisoned samples largely determines installation, while benign samples help preserve trigger conditionality, are worth further investigation.
  • The paper evaluates defenses under different access assumptions. The high-temperature probing results are particularly interesting, suggesting that conditional backdoor behaviors may become observable even without the intended trigger.
Critical weaknesses
  • SilentCall largely extends the conditional backdoor paradigm of Sleeper Agents to tool calls using straightforward paired-data SFT. GRPO offers negligible improvements, and the technical distinction from prior agent backdoor attacks, such as BadAgent, remains unclear.
  • The attack is evaluated mainly through the emission of a fixed malicious tool call with synthetic execution results. Its effectiveness in real multi-step agent tasks with dynamic tool feedback and execution failures remains unverified. Thus, the reported ASR does not reflect end-to-end attack success.
  • The Reveal metric focuses on explicit payload references in responses without independently examining reasoning traces. Since benign final responses are reused during poisoning, the reported concealment may largely reflect data construction rather than a general ability to hide malicious intent.
  • The reported 100% detection rate is evaluated against a single fixed payload, without diverse or adaptive attacks. The classifier-plus-LLM-judge approach is relatively conventional, and its effectiveness against subtle or multi-step malicious behaviors remains uncertain.
  • Experiments cover only one trigger, one payload, and two models from the same family. The weight-based audit further assumes a matched benign reference model, and its reliability across different training configurations is not established.
Questions for authors
  • What technical challenges distinguish SilentCall from existing conditional backdoors, such as Sleeper Agents and BadAgent? Beyond paired-data SFT and additional tool calls, what is the key methodological contribution?
  • Can the authors evaluate end-to-end attack success in realistic multi-step agent environments with dynamic tool feedback? How well does the attack generalize across different triggers, payloads, and tool schemas while preserving legitimate task performance?
  • Does the reported 0% Reveal rate hold when reasoning traces are inspected? Can a CoT-based monitor detect malicious intent despite normal user-facing responses?
  • How does the proposed monitor perform against unseen or adaptive attacks, particularly when malicious objectives are achieved through sequences of individually legitimate tool calls?
  • Can the authors explain why poisoned sample count determines backdoor installation while benign sample count affects trigger conditionality, beyond the reported empirical observations?
Reviewer_2Kb7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates backdoor attacks against open-weight LLM agents that use external tools. The authors introduce SilentCall, a fine-tuning-based attack in which a model learns to execute an additional malicious tool call when a predefined temporal trigger appears in the system context, while continuing to perform the legitimate task and without explicitly revealing the malicious operation in its user-facing response. The attack is evaluated on two Qwen3 model sizes using supervised fine-tuning and optional GRPO training. The authors study attack reliability, utility preservation, and the influence of poisoned and benign training data through an extensive set of experiments. The paper also investigates several detection approaches, including a cascaded runtime monitor, high-temperature probing, weight-distribution analysis, and conventional alignment benchmarks. The main findings indicate that the attack can achieve high reliability while maintaining legitimate task performance, whereas the proposed runtime monitoring approach successfully detects the evaluated malicious payload.

Critical strengths
  1. Relevant and practically important security problem The paper addresses a relevant security concern related to the adoption of open-weight LLM agents. Backdoored models that perform unauthorized operations through external tools represent a realistic and potentially serious threat, particularly when downstream users rely on model reputation and standard benchmark performance. The emphasis on monitoring actual agent actions rather than relying exclusively on user-facing responses is practically important.
  2. Extensive experimental analysis of backdoor installation The authors provide a relatively detailed investigation of the attack mechanism, including comparisons between benign and poisoned fine-tuning, experiments with GRPO, and an extensive analysis of the effects of poisoned and benign training data. The 161 training runs investigating the conditions under which the backdoor becomes reliably installed provide useful empirical observations.
  3. Evaluation of different detection strategies The paper considers several detection approaches under different assumptions about defender access. Distinguishing runtime monitoring from techniques requiring model weights or additional training resources is a useful perspective for practitioners.
Critical weaknesses
  1. Limited novelty and insufficient differentiation from existing researc My main concern is the limited novelty of the proposed attack and detection approaches. Backdoor attacks against LLM-based agents have already been extensively investigated, including attacks that manipulate intermediate agent behavior, trigger unauthorized tool execution, and preserve apparently legitimate final responses. For example, Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agents (Yang et al., 2024) introduced a framework and taxonomy of agent backdoor attacks, explicitly considering attacks that modify intermediate agent actions while preserving the correct final output. Similarly, BadAgent (Wang et al., 2024) demonstrated the implantation of backdoors into LLM agents through fine-tuning and their activation through user inputs or environmental conditions. More directly related, Data Exfiltration via Backdoored Tool Use (Zhang and Pei, 2026) already investigated data exfiltration through malicious tool calls initiated by backdoored LLM agents. In this context, SilentCall appears to be a specific implementation of an already established attack category rather than a fundamentally new attack mechanism. The use of a temporal trigger, an additional shell command, and an unchanged user-facing response does not, by itself, establish substantial methodological novelty. Although these studies are referenced in the paper, the differences are primarily discussed at a descriptive level. I would expect a more systematic comparison with existing agent backdoor taxonomies and methods, explicitly identifying the new technical challenges addressed by SilentCall and the capabilities that were not already demonstrated. Similar concerns apply to the proposed defenses. Runtime inspection of tool calls and the combination of lightweight classifiers with more advanced reasoning-based security checks are already established approaches in agent security. Existing frameworks such as LlamaFirewall and GuardAgent are acknowledged but are not included in a direct experimental comparison. Overall, the paper combines several established ideas, but I am not convinced that it sufficiently demonstrates a substantial new methodological contribution.

  2. Limited generalizability of the attack and evaluation Although the reported attack success rates are high, the experimental setting is relatively narrow. The main evaluation uses only two model sizes from the same Qwen3 family, a single temporal trigger, and one fixed malicious shell-command payload. Consequently, it remains unclear whether the reported results generalize to different model architectures, more complex triggers, alternative malicious operations, or different tool interfaces. This limitation is particularly important for the detection evaluation. The proposed runtime monitor reports 100% detection at a 1.73% false-positive rate. However, all evaluated malicious calls contain the same payload, repeated across different conversation contexts. Therefore, the reported result primarily demonstrates reliable detection of one specific malicious command rather than general detection capability against diverse backdoored tool-use behaviors. The evaluation would be considerably stronger if it included multiple independently designed malicious payloads, different types of unauthorized tool operations, and attacks using different encoding or execution strategies. Another limitation concerns the definition of attack success. The reported ASR measures whether the model emits the malicious tool call, rather than whether the corresponding operation executes successfully and achieves the intended security impact. While this is a reasonable operational metric for the controlled experiment, the distinction should be clearer when interpreting the results. The conclusion that the attack is broadly effective and that the proposed defenses can reliably detect it is therefore insufficiently supported beyond the specific evaluated setting.

  3. Insufficient validation and comparison of the proposed detection methods The evaluation of the defensive mechanisms raises additional questions regarding their novelty and practical effectiveness. The runtime monitor consists of a lightweight classifier followed by an LLM-based contextual judge. Although the reported detection and false-positive rates are promising, the paper does not provide a sufficiently comprehensive comparison with established agent security approaches. Without such comparisons, it is difficult to determine whether the proposed architecture offers a meaningful advantage over existing tool-call monitoring or policy-enforcement mechanisms. Overall, the defensive evaluation would benefit from stronger baselines, a broader range of attacks, and more rigorous validation under realistic defender assumptions.

Questions for authors
  1. Novelty relative to existing agent backdoor research Could the authors clarify the fundamental technical novelty of SilentCall compared to Yang et al. (2024), Wang et al. (2024), and Zhang and Pei (2026)? In particular, how does SilentCall differ from previously studied backdoors that manipulate intermediate agent actions while preserving legitimate final responses? A structured comparison of threat models, triggers, malicious operations, concealment mechanisms, and demonstrated capabilities would help clarify the contribution.

  2. Generalizability of the attack Can the authors demonstrate that the observed attack reliability extends beyond the single evaluated payload and temporal trigger? It would be useful to evaluate different types of malicious tool calls, trigger mechanisms, and model families to establish whether the findings are general properties of agentic backdoors or specific to the selected experimental configuration.

  3. Evaluation of the runtime detection approach How would the proposed runtime monitor perform against a diverse set of malicious tool calls rather than the same fixed payload repeated across different conversations? Additionally, can the authors provide direct experimental comparisons with existing agent security frameworks, such as LlamaFirewall or GuardAgent, under comparable conditions? This would help establish whether the proposed defense provides a meaningful improvement over existing approaches.

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

Related papers

Discussion (0)

Sign in to comment.