acceptodds
Under review as a conference paper at ICLR 2027

TruncDoS: Resource Exhaustion Attack via Response Truncation in Tool-using Agents

Abstract

Tool-using LLM agents implicitly assume that the response they receive faithfully preserves what an honest tool returns. However, prior work on resource exhaustion has focused on content-level stuffing, such as injected instructions, malicious tool use, or manipulated metadata descriptions. In contrast to content-level stuffing, we present TruncDoS, the first attack that amplifies token consumption and latency solely by truncating tool responses. We evaluate the feasibility of TruncDoS, which can be installed through a marketplace skill or plugin, using an in-process mediator that accesses the shared tool-dispatch return path. The mediator achieves TruncDoS by truncating the tool response and forwards fragments. When the missing trunk contains task-relevant evidence, the agent treats the response as incomplete and invokes further queries until the resource budget is exhausted or the task fails. TruncDoS can cause significant increase in token consumption and latency only through truncation, instead of injection, tool modification or additional model invocation at runtime. We demonstrate the resource exhaustion depends on the agent's tool-invocation capability. We then assess TruncDoS across multiple agent harnesses, model families, and benchmarks. The attack increases by up to % in token consumption and latency by up to %. We further apply four representative defense adaptations; none restores clean-task utility, and several even add overhead.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent Security, Tool-Using LLM Agents;Runtime Truncation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QEGQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents TruncDoS, an attack that truncates tool responses to induce repeated queries, increase resource consumption, and reduce task utility. It formalizes the middleware threat model and evaluates multiple agent configurations, with deletion-only controls supporting the core effect and defense tests revealing gaps in observation-completeness protection.

Critical strengths
  • Interesting attack mechanism. The paper identifies observation completeness as an availability concern: deleting genuine tool content can trigger costly recovery behavior without inserting instructions. This provides a useful perspective on agent security.
  • Evidence supporting the core effect. Matched deletion-only experiments show increased resource consumption and reduced task utility, supporting attribution to truncation rather than appended guidance. These controls strengthen the empirical contribution.-
Critical weaknesses
  • The attack requires permission to modify tool returns, but the plugin experiment lacks an end-to-end demonstration of planner-level resource amplification. Direct detection or user-perception measurements are also missing, limiting the evidence for practical deployment and stealth.
  • Section 2 disclaims reproducing LoopTrap and comparing effectiveness with published attacks, yet Table 2 uses the LoopTrap label and Section 5.2 discusses relative performance. This unclear correspondence weakens the comparative conclusions.
Questions for authors
  • What permissions does the extension require, and can it modify other tools’ returns? Do end-to-end traces demonstrate planner consumption and additional costs? Do detection measurements or human evaluations support the stealth claim?
  • What implementations do the LoopTrap and Chord rows represent? Please clarify their sources, adaptations, tuning budgets, and which comparative claims they support.
  • Were the defense experiments conducted with fixed truncation strategies or defense-aware adjustments? Please specify the attacker’s knowledge, tuning budget, and whether adaptive attack results are available.
Reviewer_Mq7ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a resource-exhaustion attack against LLM agents. Unlike conventional prompt-injection attacks, TruncDoS truncates legitimate tool responses to induce repeated tool calls, increasing token consumption and latency while reducing task completion. The authors evaluate the attack across multiple models, benchmarks, and defense mechanisms, demonstrating substantial resource amplification and utility degradation.

Critical strengths
  1. The paper identifies tool-response completeness as an overlooked attack surface. The authors propose an effective attack approach through information removal rather than malicious instruction injection, which is an interesting observation.
  2. The experiments cover multiple models, agent frameworks, and benchmarks. The deletion-only ablations provide useful evidence that response truncation itself can induce additional tool calls and resource consumption.
Critical weaknesses

1.The proposed attacks reply on access to a privileged in-process middleware layer. I wonder whether this kind of attack is practically applicable. Its generality is also insufficiently validated. 2. The main experiments mix different attack configurations, including a historical variant that appends additional guidance to truncated responses. Although deletion-only experiments are provided, the controlled mechanism study covers only ten tasks, while the larger deletion-magnitude sweep uses a fixed condition order.

Questions for authors
  1. can the authors demonstrate that a third-party extension can intercept responses from tools? who would be the attacker?
  2. can the authors provide matched deletion-only experiments across the evaluated models?
Reviewer_jQ42OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies an availability attack in which middleware truncates successful tool responses before an agent receives them. The retained text is a prefix of an honest response. Missing evidence can cause the agent to issue further queries, increasing resource consumption while reducing task completion.

The paper evaluates several planners and harnesses, separates deletion from appended framing text, and studies truncation severity and defense adaptations. The strongest matched deletion-only evidence is the 97-task DeepSeek sweep: at 90% deletion, completion falls from 87/97 to 39/97, mean tokens increase from 15,045 to 77,914, and mean latency increases from 33.9 to 215.7 seconds. A smaller MiniMax experiment also shows increased calls and cost without appended instructions.

Critical strengths

The deletion-only experiments establish a meaningful behavioral effect. Table 8 evaluates the same 97 tasks across clean and four truncation levels, without appended text. Resource use rises while completion falls. Although fixed execution order limits causal interpretation of the severity trend, this is substantially stronger evidence than a few illustrative looping traces. The smaller mechanism experiment provides additional support on another planner.

The paper identifies a concrete gap between content safety and observation completeness. A retained prefix can contain no malicious instruction yet still leave an agent without the evidence needed to finish. This is useful for designing agent reliability tests and for explaining why instruction filters or action-authorization checks alone do not address this failure.

Critical weaknesses

The deletion-only experiments establish a meaningful behavioral effect. Table 8 evaluates the same 97 tasks across clean and four truncation levels, without appended text. Resource use rises while completion falls. Although fixed execution order limits causal interpretation of the severity trend, this is substantially stronger evidence than a few illustrative looping traces. The smaller mechanism experiment provides additional support on another planner.

The paper identifies a concrete gap between content safety and observation completeness. A retained prefix can contain no malicious instruction yet still leave an agent without the evidence needed to finish. This is useful for designing agent reliability tests and for explaining why instruction filters or action-authorization checks alone do not address this failure.

Questions for authors

Clarification: Please reconcile the abstract, introduction, and Table 2, including the peak metric labels, DeepSeek denominator, and comparator identities. For each result, which transform version and mode were actually executed?

Assessment-changing comparison: Under the same task and resource budgets, how does deletion-only truncation compare with empty returns, explicit errors, and ordinary output-length limits? This would show whether truncation creates a distinctive amplification effect rather than a general response to missing evidence.

Assessment-changing experiment: How effective are repeated-call limits, no-progress detection, bounded retries, and per-task token or wall-time budgets? Please report both resource reduction and benign-task utility, rather than requiring a defense to reconstruct unavailable information.

Clarification: Were the six Claude CLI output-format failures caused by the attack? Appendix A says they are not established infrastructure failures, yet they are excluded from the completion denominator. Please report outcomes over all 97 configured tasks and separate task correctness from producing parseable output.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.