- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents Clawdrain, a resource-exhaustion attack delivered through a Trojanized third-party OpenClaw skill. The attack implements a Segmented Verification Process (SVP) that withholds an otherwise legitimate payload until an agent completes several validation stages. Each stage requires the agent to construct and submit a calibration sequence to a companion validator, which returns progress, repair, or terminal signals. The goal is to increase victim-funded model and tool computation while still returning a normal final answer.
The paper's main empirical argument is that deployed-agent resource cost cannot be inferred from output length or nominal tool-call count alone. In a complete OpenClaw runtime, accumulated context may be repeatedly re-ingested, and agents may respond to the malicious workflow through direct generation, scripting, tool coordination, retries, or failure recovery. The authors additionally introduce SVP+N, which provides fresh per-stage checkpoint values to prevent exact reuse of an earlier calibration sequence.
The evaluation uses real OpenClaw deployments and commercial API backends for Gemini 2.5 Pro, DeepSeek V4 Flash, and Claude Sonnet 4.6. With 50 runs per condition, plain SVP yields median cumulative-input amplification from approximately 2.7× to 16.4× over matched ordinary requests. The paper also documents qualitatively different backend behaviors and analyzes costly failed executions that consume substantial resources despite not completing the requested task.
I find the deployment-grounded measurements useful, particularly the observation that resource costs can migrate from visible generation to cumulative context processing, coordination, and recovery. My main concerns are novelty relative to closely related agent resource-exhaustion attacks and the lack of a matched benign multi-stage baseline or causal ablations. The current comparison establishes that a malicious skill that deliberately induces additional multi-turn work is more expensive than a normal request, but it is less clear how much of the reported amplification reflects a specifically exploitable runtime pathology rather than the expected cost of executing a more complex workflow.
Critical strengths
The strongest aspect of the paper is its deployment-grounded evaluation. Rather than evaluating an isolated tool loop, the authors execute Clawdrain in a complete OpenClaw runtime with real model APIs, persistent context, general-purpose tools, and actual provider usage accounting.
The paper also provides a useful systems-level perspective on resource exhaustion. For Gemini, input amplification is much larger than output amplification, demonstrating that economic impact need not manifest as visibly long model responses. The cached/uncached accounting further shows that repeated context remains relevant even when providers optimize repeated prefixes.
The cross-model behavioral analysis is informative. Gemini transitions from direct sequence generation to scripting, DeepSeek scripts proactively, and Claude exhibits substantial trial-and-error and recovery behavior. These results show that capable tool use can reduce one component of an attack without necessarily eliminating cumulative execution cost.
I also appreciate the separation of resource consumption from attack completion and final-answer correctness. The Claude failure cases illustrate why conventional ASR can be misleading for denial-of-service attacks: a run may fail to return the desired answer while already imposing substantial cost. The threat model is plausible for an open third-party skill ecosystem, and the paper is careful not to equate answer preservation with measured stealth. The authors additionally retain timed-out and compacted executions rather than discarding inconvenient traces.
Finally, performing 50 fresh-session trials per condition is substantially stronger than a small proof-of-concept evaluation.
Critical weaknesses
My primary concern is that the central experiment lacks a matched benign multi-stage baseline. The Base condition disables SVP entirely. Clawdrain therefore introduces additional stages, tool calls, and large arguments by design, so significantly higher resource consumption is expected even without any unusual runtime effect. A benign workflow matched for stage count and/or total payload size would help isolate how much amplification is specifically due to context re-ingestion, repair, and malicious workflow structure. Relatedly, the paper provides a mechanistic interpretation involving persistent context, sequence generation, relay state, recovery, and tool coordination, but explicitly states that the measurements capture their combined effect and do not isolate the individual contributions. Controlled ablations would make the central systems claim considerably stronger.
The novelty over prior resource-exhaustion work is also somewhat incremental. The related work already includes malicious MCP/tool attacks that generate costly multi-turn trajectories, including attacks preserving a correct final response. I see the main contribution as the production-runtime measurement and analysis of agent adaptation, rather than the basic staged exhaustion attack itself.
The quantitative reporting does not fully exploit the 50 trials per condition. The main results emphasize medians and a binary result marker but do not prominently report dispersion, timeout/compaction frequencies, or answer-success rates. Because the paper emphasizes model-dependent adaptive behavior, these distributions seem important.
The SVP operating points additionally vary sequence length, stage count, and instruction detail simultaneously, so they cannot support clean conclusions about attack scaling with any individual parameter.
The proposed runtime defenses—cumulative per-skill budgets and retry/loop circuit breakers—are sensible but are not implemented. A simple quantitative defense evaluation would substantially strengthen the paper and clarify whether the attack exposes a fundamental challenge or primarily a missing accounting mechanism in the current OpenClaw runtime.
Finally, external validity is limited to one agent runtime with general-purpose shell/Python access. Different context-management policies, tool permissions, spending caps, or skill sandboxes may substantially change the observed amplification.
Questions for authors
-
Can the authors add a benign multi-stage workflow matched to SVP in stage count, tool-call count, or total payload volume, to separate attack-specific amplification from the ordinary cost of additional work?
-
Can the authors isolate the contribution of context re-ingestion, turn count, sequence length, repair, and recovery through controlled ablations?
-
What are the IQR/confidence intervals and per-condition rates of successful task completion, timeout, context compaction, and final-answer correctness across the 50 trials?
-
How effective are the proposed cumulative per-skill budgets and loop/retry circuit breakers, and what utility degradation do they cause for legitimate multi-stage skills?
-
Given the close relationship to prior MCP/tool-based multi-turn resource-exhaustion attacks, can the authors more sharply articulate which security phenomenon is newly established by Clawdrain beyond deployment realism?