- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies denial-of-service attacks against LLM-based agent guardrails. Attackers embed text that mimics safety-review procedures into webpages, repository files, or tool outputs, to enumerate risks. Two optimization approaches are proposed: beam search using an LLM proposer, and structural mutations guided by internal model signals.
Critical strengths
- This is an important and well-motivated problem. The paper extends guardrail evaluation beyond verdict correctness to timely completion, which matters for practical deployments.
- Exploiting the structure of safety analysis is well suited to guardrails and more compelling than injecting unrelated computationally expensive tasks.
- Experiments span multiple models, benchmarks, guardrail templates, and agent workflows, providing evidence of transferability and downstream impact.
Critical weaknesses
- The attack path in a key deployment is insufficiently documented. The Claude Code experiment needs to explain precisely how repository guidance reaches the safety classifier and provide sanitized request traces to establish consistency with the threat model.
- Limited algorithmic novelty and incomplete comparisons. Beam search, strategy banks, and structural mutations are relatively conventional. Comparisons against simple handcrafted templates, random search, and appropriately adapted prior attacks under equal budgets are missing.
- Results emphasize means and maximum amplification without confidence intervals, loop-trigger rates, or truncation frequencies, making attack reliability difficult to assess.
- Some typos should be revised, such as "a LLM" -> an LLM, "Table 11 report"-> Table 11 reports, etc.
Questions for authors
Can you demonstrate an action that is correctly blocked without the attack but executes after an attack-induced timeout? This would distinguish an actual safety bypass from the execution of an otherwise authorized action without review.
Response to fKTq on deployment evidence, comparisons, and reliability
Thank you for the constructive feedback. We address each point below.
1. How repository guidance reaches the Claude Code classifier.
The attacker poisons repository-root
CLAUDE.mdin a cloned repo; a user runs the stock Claude Code CLI to write.claude/settings.local.json. While Auto Mode stripsReadoutputs from<transcript>and auto-allows routine edits, the payload reaches the classifier natively:CLAUDE.mdverbatim in<user_claude_md>insidemessages[0]of every classifier call..claude/settings.local.jsontriggers.claude/settings*.json(Self-Modification), invoking Stage 1 (max_tokens=64, stop</block>). After<block>yes</block>, Stage 2 (max_tokens=4096) receives<user_claude_md>+<transcript>and follows the injected schema in<thinking>, raising Stage 2 tokens by4.9–7.3×and latency by6.8–9.9×over benign Stage 2 checks (10.1–14.6×end-to-end; Table 6). Responses are captured via a read-only proxy; path and sanitized traces are in Secs. 3.2, 5.1, and App. A.5.1.2. Algorithmic novelty and baseline comparisons.
Beam search, strategy banks, and slot mutations are standard primitives. Our novelty lies in uncovering a new security surface—turning task-confined guardrails into an availability bottleneck via schema mimicry—and characterizing its mechanism (
9.6×header-attention cycling and low-entropy overfitting; Sec. 3.3.1), informing our zero-LLM search (Inst. II, Eqs. 4–5) and defense (Sec. 6).All three baselines are in the paper (Sec. 4.2, App. B.4):
Iter 0; Tables 3, 14, 18) and minimal schemas (Table 16) yield≤2.6×extension without loops.−26%) or greedyk=1search (−88%) sharply drops fitness.≤1.20×length,≤3.7×GPU time; Table 2), as guardrails ignore unrelated decoy tasks.3. Reliability, loop-trigger rates, and truncation frequencies.
Table 1 optimizes on
2cases and tests zero-shot on20held-out cases ×5runs (100runs/cell) (Sec. 4.1, App. B.1). Because responses are bimodal (Sec. 4.2)—short non-loop expansion vs. cap-approaching loops—a single Gaussian mean±CI straddles both regimes; we report ranges plus trigger and cap-hit rates:>3repeats/500chars; App. B.1),20/60Table 1 cells loop (*), and40/60still reach13.2–43.8×. Per run/task:8/10TS-Guard-8Btraces loop (Sec. 6);7/10Claude-Haiku-4.5and9/10GPT-5.5runs exceed20×latency (App. B.3);18/20README and13/20config tasks on OpenHands exceed3×(Table 11).16,384tokens (28,000in LangGraph, hit by>82K-char peaks). In Claude Code (4,096-token cap; App. A.5.1),2/12seed and1/5finalist runs hit the cap vs.0/6benign Stage 2 runs. In WebGuard (Table 9), timeouts hit0baseline steps vs.17(allow) and34/55(61.8%,block).4. Matched demonstration: blocked action executing after timeout.
Yes. We show webagent example: fixing one
HIGH-risk purchase—click(33)(Place Order) on/checkoutwith pre-filled card details inBrowserGym + WebGuard(gpt-4o-mini, identical prompt,0API errors across 15 runs)—we varied only whether the page embeds frozenWEBGUARD_PAYLOAD_V1:30sfail-open*
2/5hit16,384tokens (165.7s) withoutrisky_level(0/5executed; even a120scutoff truncates before a verdict).Clean runs block
Place Order5/5times (4.5s), and the payload never convinces WebGuard purchasing is safe (0/5allowed without fail-open; the2/5non-looping30sruns still block in4.6–11.0s). Yet in3/5runs, schema-filling streams up to15,001chars past30swithoutrisky_level, so fail-open executesclick(33)to/order-confirmation(App. A.5.3).5. Presentation.
We fixed
"an LLM"and"Table 11 reports"in the revised paper.