acceptodds
Under review as a conference paper at ICLR 2027

Unwilling, or Unable? Identifying the Safety Cost of Tool Interfaces in LLM Agents

Abstract

Agent-safety benchmarks typically use refusal rate, but a model that declines because a request is harmful and one that declines because it has no way to act look identical in text. We show that this confound can materially change conclusions for several evaluated models. We use AgentHarm’s built-in benign twins which are matched, harmless tasks that share each harmful item’s structure and grader and construct a capability-adjusted estimator: a difference-in-differences that subtracts the twin’s refusal change, plus an item-level rate conditioning on a demonstrated capability match. The core estimators, audit framework, exclusion rules, and minimum-sample requirements were documented before the later hosted and local evaluation tiers; subsequent refinements are recorded in a timestamped deviation log, and reported results are generated from the released analysis pipeline rather than transcribed. We evaluate 16 hosted models and 8 locally served configurations spanning 10 organisations overall. On the 14 hosted models that support both interfaces, the mean measured effect moves from -0.319 to -0.188, 3 point estimates reverse sign, and 6 lose significance after correction. When both sides expose an executable action space, refusal-to-action conversion ranges from 3.5% to 77.8% across models. In separate multi-turn runs, 34.1% of GPT-4o’s demonstrated refusals become actions (20.0% fully completed), 58.9% of Llama-3.3-70B’s (21.1%), 62.5% of Mistral-Large-3’s (21.4%), 38.7% of Qwen3-32B’s (16.1%), and 6.8% of DeepSeek-V3.2’s (1.7%). On the three tested OpenAI models, the capability-adjusted graded-harm change is negative: raw harmful completion rises, but by less than completion on the matched benign controls. On the tested openweight configurations, the largest behavioural shift occurs when an executable action space first appears, while hosted native-channel effects are model-dependent. Harm remains linearly decodable at the decision token within framing on the Qwen-family states we instrument. A linear probe guard substantially reduces completed harm but exhibits calibration drift, and a dedicated decline task action restores refusal under a matched placebo on Qwen3-8B and hosted GPT-4o, while worsening harm on Gemma-4-E4B.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, AI safety, tool use, function calling, refusal behaviour, agent evaluation, capability confounding, mechanistic interpretability

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

The paper argues that refusal-based evaluations confound safety with inability to act. Using AgentHarm’s matched harmful/benign tasks, it proposes a capability-adjusted DiD that subtracts the benign refusal change from the harmful refusal change, together with conversion rates conditioned on demonstrated capability. Across 24 configurations, the average hosted-model effect decreases from −0.319 to −0.188 after adjustment; several estimates lose significance. The paper also studies where the interface effect occurs, whether harmfulness remains linearly decodable, and two diagnostic interventions.

Critical strengths
  1. The paper identifies an important confound in agent-safety evaluation. The matched controls and capability-conditioned conversion rates are more informative than a simple comparison of refusal rates.

  2. The evaluation includes broad model coverage, clustered inference, multiple-testing correction, judge validation, parser and cell-health audits, and several sensitivity analyses.

Critical weaknesses

The central interpretation depends on a strong parallel-response assumption: interface changes must affect harmful and benign twins similarly except through safety. This is not directly validated. Indeed, many benign refusals are themselves safety-framed, and the conservative sensitivity materially changes some estimates. The evidence supports the narrower claim that naive effects are overstated, but not necessarily that the residual identifies the interface’s safety cost.

The contribution is incremental. Capability confounding, matched benign controls, text/action divergence, and tool-interface effects are already established. The main novelty is their combination in one estimator, which is evaluated primarily on AgentHarm. The probing and intervention results are limited or mixed and do not establish a general mechanism.

Questions for authors

Can the authors provide stronger empirical support or placebo tests for the parallel-response assumption?

How do the headline results change if all records with conflicting reruns are excluded, rather than resolved using the last-write rule?

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

This paper takes a fresh look at refusal rates in AI agents, pointing out that comparing agents with and without tools mixes up harm avoidance and simple inability to act. By introducing AgentHarm's matched benign twins, the authors propose a new difference-in-differences approach and compare refusal-to-action conversion rates, but only for cases where the agent has already demonstrated it's capable (using twin-conditioned and forced-action witnesses).

Critical strengths
  1. the usual comparison actually exaggerates the safety cost of interfaces. is carefully demonstrated with a proper control (the benign twin), which the field hasn't used before. The authors are open about their finding: it's an overstatement, not an outright reversal, as Table 7 shows.
  2. the stats are handled with care: clustering at the level of 44 original behaviors, a Holm correction for multiple comparisons, audit gates to catch weird outputs and attrition, a parser for five tool-call syntaxes, judging validation with both human raters and several judges, plus a log of deviations.
Critical weaknesses

The parallel-response assumption is stronger than the authors admit, and the bias runs in a predictable direction. When agents decline to act, that's another reason for refusal besides harm, yet this only shows up on items not already refused for harm. So, harmful items already refused can't stack up additional refusals for inability, making their drop from this source smaller than the benign twin’s. Subtracting the entire benign drop ends up over-correcting, which pushes estimates toward zero or even positive numbers, exactly where those three sign reversals pop up. The conservative sensitivity check in Section 5 tackles something else (benign refusals that are actually safety-driven). Also, the additive probability-scale DiD doesn’t have a logistic-link equivalent, and twins aren't matched for difficulty, tool names match for 143 out of 176 pairs.

For the conversion numbers, there’s no null reference point. When they condition on “refused in prose” and check action under native tools, it picks up noise or sensitivity to how prompts are presented, since refusals close to the decision boundary flip easily. Sure, they have a prose2 cell and benign-twin refusals, but they don’t run conversion for a placebo contrast (like prose versus prose2 or twin items refused in base). Their checks only repeat four GPT-4o cells. The eligible denominators are small, intervals are wide, and the single-turn and multi-turn group sizes don’t match. Sections 7 and 8 (covering probes on five Qwen-family states and decline_task) don't connect to the estimator and, honestly, are inconsistent, it helps some models, worsens others, so overall measurement gets diluted. Table 6 shows multi-turn DiD values (-0.045, -0.341) that aren’t defined in the main text and don’t clearly line up with Table 4.

Questions for authors

What’s the conversion rate in a placebo contrast, and does the DiD sign analysis hold up if you use a bounds- or ceiling-corrected approach?

Can you provide a risk ratio or logistic (link-scale) analysis, at least for OpenAI and Bedrock panels?

How does the conclusion shift if you stick to the 143 exact-tool pairs for all models?

Which metric do the multi-turn DiD values in Table 6 refer to, and why don’t they match Table 4?

Is the interface effect stable if you try a second matched interactive suite, even a small one?

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

This paper's main goal is to remove the confound of whether an LLM is able to / good at using tools from whether an LLM agent refused to do a harmful action. Their approach is to take benign versions of AgentHarm tasks and no-tool-using versions, and compute a diff of diffs to see whether agents are truly safe for agentic tasks.

Critical strengths
  • I find the paper's main idea very interesting: separating inability to do something because of lack of agentic capabilities vs. actual safeguards activated via refusal to act on something the agent could have.
  • The use of AgentHarm scenarios is a good idea, and from what I can understand, the operationaization of their approach makes sense.
Critical weaknesses

My main issue is that the paper is largely written in a way that makes it quite incomprehensible (I feel like I'm recognizing Claude's writing style, which I find to be incredibly unclear, but I could be wrong). This makes it really hard to digest what the paper is actually doing, what the results are referring to, etc. Largely, this is because the paper does not follow the standard structure of Approach -> Results -> Discussion/Conclusion, reads more like a blogpost, doesn't signpost enough, tends to use various ill-defined or overly metaphorical words (e.g., "surface", "moves", "shed"), lack of consistent terminology, doesn't have good running examples, switches between experimental setups, does not properly contain takeaways for results, and other issues.

It's also missing a related work in the main body of the paper, and in the appendix' related works there are missing some works that tackle very similar research questions. For example, https://arxiv.org/abs/2409.16427 examined the effect of tool-access on agentic safety. Other works include https://arxiv.org/abs/2507.06134, https://arxiv.org/abs/2602.13379, etc.

I note that my assessment of the paper is negative purely due to the poor writing which I know the authors can easily improve, but likely not within the review window. If ICLR / the ML community had revise-and-resubmit, I would have been happy to re-review the paper in a better written format, because I believe the contribution could be very interesting!

Though I believe GenAI could likely help with improving the writing (asking it to make the paper clearer or less LLM-y can help), some guides for better writing (at the risk of sounding condescending):

Questions for authors

L211: What does "disk-cached GPT-4o-mini" mean? How can a proprietary model be disk-cached?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.