acceptodds
Under review as a conference paper at ICLR 2027

SkillDOS: Token Amplification Attacks via Poisoned Skills in LLM Agents

Abstract

Skills enable LLM-based agents to follow reusable workflows, invoke external tools, and complete complex tasks, but they also introduce a new attack surface. Existing research on agent-skill security has primarily focused on prompt injection, data exfiltration, and unauthorized actions, while resource-exhaustion attacks through poisoned skills remain comparatively underexplored. We introduce SkillDOS (Skill-based Denial of Service), an automated framework that amplifies an agent's output-token consumption through poisoned skills. Rather than causing explicit failures or non-terminating execution, SkillDOS induces additional bounded, task-relevant work while preserving the final user-visible answer. SkillDOS decomposes a poisoned skill into an Inducer, which promotes adoption of an injected workflow, and an Amplifier, which scales workload through repetition, grounded work-unit expansion, and per-unit elaboration. A one-way two-stage optimizer first refines the Inducer until execution evidence confirms a distinctive Amplifier mechanism, then freezes the Inducer and searches Amplifier configurations for greater output-token amplification under an answer-preservation constraint. We evaluate SkillDOS on 100 real-world skills across three agent frameworks and four backend models using a sandboxed paired-baseline protocol. The results demonstrate substantial output-token amplification across diverse agent-model configurations while maintaining high task preservation. These findings reveal a practical and underexplored resource-exhaustion risk in skill-enabled LLM agents.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLMs, Autonomous Agents, Skill Injection, Denial of Service (DoS), Token Amplification, AI Security.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TwWiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

SkillDOS is a token-amplification (“denial of wallet”) attack on skill-enabled agents. A poisoned skill splits into an Inducer (gets the workflow adopted) and an Amplifier (scales bounded, task-relevant work), optimized in two stages under an answer-preservation constraint. Across 100 skills, three frameworks, and four models, it reports mean token-amplification ratios of 2.65–13.96× at 95–100% task preservation.

Critical strengths
  1. Genuinely new attack dimension. Prior skill-security work targets payload execution or exfiltration; resource amplification that preserves the answer while inflating cost is a distinct and under-studied surface. The denial-of-wallet framing is timely for usage-metered agents and well-motivated.
  2. Clean formalization. The TAR ratio, the baseline-relative preservation judge, and the one-way two-stage optimizer (freeze Inducer once activation is verified, then search Amplifier genomes) are precisely specified. The activation-vs-preservation separation is a sensible design that avoids conflating adoption with amplification.
  3. Reasonable breadth and controls. Evaluation spans three agent frameworks and four backends with a paired sandboxed baseline, plus a SkillJect comparison, cross-model transfer, a stagewise ablation, and a scanner-evasion test. The ablation cleanly attributes most gain to Stage 1.
Critical weaknesses
  1. Means are driven by outliers. Medians are modest (2.05–5.88×) while means reach 13.96× and a single case hits 558×. Leading with means materially overstates typical amplification; a 2–3× median is a far less alarming result than the abstract implies.
  2. Single-baseline noise. The clean baseline is run once per case (acknowledged), so model stochasticity is uncontrolled. Given that small TARs (~2×) sit near plausible run-to-run variance, part of the reported amplification may be noise rather than attack effect.
  3. Weak evasion undercuts the threat. Existing scanners already detect 73.3% of poisoned skills, so only ~27% evade. Combined with an attack-only contribution and no proposed defense, the “practical risk” claim is softer than stated.
  4. Circular preservation metric. Preservation is an LLM judge that ignores length and verbosity — but inflated length is the attack. A 10× longer “equivalent” answer counts as fully preserved, so the metric cannot distinguish harmful padding from legitimately added value.
Questions for authors

see weakness

Reviewer_Q4q2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work, SkillDOS poisons an agent’s skill instructions to make it do extra work and burn more tokens, while keeping the final answer largely intact. It uses an LLM to generate and refine the payload, first getting the agent to follow the added workflow and then increasing its workload. The results show token amplification across several frameworks and models, but comparisons with existing attacks are limited.

Critical strengths

The broad evaluation across 100 skills, three agent frameworks, and four models, showing that token amplification isn’t limited to one setup.

The method also checks whether the agent actually performs the added work and tracks answer preservation, rather than treating any longer or broken run as a successful attack.

Critical weaknesses

I feel the baseline comparison is too narrow. Table 2 compares only against SkillJect on one framework–model configuration. Since SkillJect was not designed to maximize token consumption while preserving answers, outperforming it does not establish a strong advantage over resource-consumption attacks.Some related work is missing from the comparison, e.g. [1]. Is there any difference? The paper needs to clarify what its formulation and optimization add beyond these approaches.

The ablations don’t really show why the optimization design is needed. Comparing initialization, Stage 1, and the full method does show gains from extra optimization, but it also gives later stages more trials. Equal-budget comparisons against simple workload instructions, random search, or joint optimization would be more convincing.

The evaluation has some reliability limits. Each case uses a single clean run, so normal variation can affect amplification ratios. Answer preservation is judged by an LLM comparing final responses, which doesn’t fully verify task correctness or the artifacts produced.

The practical impact looks a bit weaker than the headline numbers imply. Some averages are likely being pulled up by extreme cases, and the paper doesn’t directly measure monetary cost, latency, or service disruption. So the evidence is stronger for token amplification than for actual denial of service.

This attack seems like prompt injection, just aimed at token amplification. Overall, that doesn’t feel like a particularly interesting new direction. It might also be worth comparing it with some prompt injection methods.

[1] Clawdrain: Exploiting Tool-Calling Chains for Stealthy Token Exhaustion in OpenClaw Agents.

Questions for authors

see above

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

This paper introduces Skill DOS, an automated framework that significantly increases the consumption of output tokens by poisoning skills. Skill DOS does not lead to explicit failure or non-termination; instead, it induces additional limited and task-related computational work while maintaining the visibility of the final user's answer. Skill DOS decomposes the poisoning skill into two components: Inducer, which is used to promote the adoption of the injection workflow; Amplifier, which enhances the workload through repeated execution, based on unit-based expansion, and unitization refinement. A one-way two-stage optimizer first refines the inducer until the execution evidence confirms the existence of a unique amplification mechanism; then it freezes the inducer and, while maintaining the answer, searches for an amplifier configuration that can achieve a higher output token amplification effect.

Critical strengths
  • It separates the induction execution from the workload amplification, and confirms whether the amplification mechanism is truly triggered through the execution trajectory, making the attack process clearer and facilitating the analysis of the roles of each stage.
  • The experimental coverage is extensive. The assessment includes 100 skills, 3 agent frameworks, and 4 types of models, and additional cross-model transfer experiments after freezing malicious skills have been conducted, providing support for the existence of risks under different configurations.
Critical weaknesses
  • I don't think the controlled experiments fully prove the advantages of the two-stage design. The main comparison method is that SkillJect has different attack targets; the existing stage ablation also increases the number of searches, lacking direct optimization or joint optimization comparisons under the same call budget.

  • The core evaluation indicators are still incomplete. The paper defines attack success as "doubling by at least 10 times and the answer remaining unchanged", but the main table does not report this joint success rate. Additionally, answer retention is only judged by comparing the final text answer and cannot fully prove that the actual deliverables such as code and files are still correct. At the same time, the verification of actual resource damage and budget defense is insufficient. The paper mainly measures the amplification of output tokens, does not assess actual costs, execution delays, nor tests cumulative tokens or call limits. The existing scanner experiments cannot explain how much damage would be caused when the budget is controlled.

  • The text in Figures 1 and 2 is too small, and Figure 2 simultaneously contains many modules, formulas, and icons, making it difficult to read after scaling down. The hierarchy of key processes and auxiliary details is not clear enough, affecting the understanding of the method.

Questions for authors
  • Report the attack success rate as formally defined, review the actual task outputs, and prioritize supplementing the experiments comparing budget optimization and cumulative token limit.

  • Increase the font size, reduce decorations and repetitive text, highlight the two-stage transition and freeze relationship; accurately express the conclusion that is supported only by text comparison as "Final Answer Maintained".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.