acceptodds
Under review as a conference paper at ICLR 2027

Benchmarking Financial Agent Safety: Scenarios and Executable Compliance

Abstract

Financial agents must complete useful tasks while respecting permissions and constraints that depend on business state. We introduce an asset-management safety benchmark organized around three deployment settings: customer service, internal assistance, and autonomous production. They differ in requester authority, permitted actions, and requirements for human review. Within thirteen business subscenarios, 877 executable cases tie safety requirements to recorded permissions, business state, and explicit rules. Adversarial and non-adversarial cases test both induced violations and failures during routine requests. Checkers inspect actions, intermediate events, and state to score violations, delivery, and safe completion; selected tasks also test safety under continuation. We evaluate nine models across 144 model-agent-defense configurations. Aggregate scores conceal business-specific weaknesses; agent implementations can improve or worsen a model's safety, as when Hermes lowers Qwen3.6-27B's adversarial violation from 24.6% to 12.8% but raises Qwen3.8-27B's from 7.9% to 19.4%; and first-turn safety can fail under continuation. Added defenses provide different benefits across attack types and can reduce legitimate task completion. A controlled study further separates claimed authorization from legitimate record changes in follow-up turns. Code and cases are available at https://anonymous.4open.science/r/fin-agent-safety.

Primary area:
datasets and benchmarks
Keywords:
Financial Agent, safety, benchmark

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_L59YOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper builds a safety benchmark for financial agents, with 877 executable cases covering customer service, internal assistance, and autonomous asset management. It looks at what the agent actually does with its tools, the current business state, and the relevant permission rules to check both whether it breaks a rule and whether it completes a legitimate task. Across different models, agent frameworks, and defenses, the results vary quite a bit by setting. One interesting finding is that an agent may refuse an unsafe action at first but go on to violate the rule later in the conversation.

Critical strengths
  1. I appreciate the authors’ effort to study the safety of financial agents in practical business settings. I find the proposed benchmark a valuable contribution, particularly because it considers both safety and task utility. An agent that avoids violations by refusing every request would have limited practical value, and the distinction between violations, task completion, and safe completion captures this concern well.

  2. Another strength is the use of actual execution traces and business states to evaluate safety, together with coverage of different deployment settings and financial scenarios. The experiments also offer useful insights into how agent frameworks and added defenses affect both safety and legitimate task completion, and how initially safe behavior can break down in later interactions.

Critical weaknesses
  1. Although the paper evaluates nine models, they come from only three model families. Including other major families, such as Claude and Gemini, would help show whether the findings generalize beyond the tested models.

  2. The threat model could be explained more clearly: who is the attacker, what can they control, and can they adapt based on the agent’s feedback? It would also help to distinguish direct requests for prohibited actions from indirect prompt injection, and provide a breakdown of attack strategies and input surfaces. Broad business coverage alone does not establish attack diversity.

  3. Attacks and defenses are often a cat-and-mouse game: defenses are designed for particular threats and need the right information to detect them. Here, many violations depend on permissions or financial limits, which the tested defenses do not explicitly check. Their limited effectiveness may therefore reflect this mismatch. Could the authors include a defense that checks business rules and current state before execution?

  4. In the record-update study, the underlying state changes without notifying the agent. It is unclear whether agents are told that records may change and should be checked again. Adding conditions with an explicit re-check instruction or a trusted update notification would help distinguish failures to seek fresh information from failures to use it correctly.

Questions for authors
  1. I found the comparison with recent finance-specific agent benchmarks somewhat limited. In particular, ReguSim (https://arxiv.org/abs/2608.19974) also evaluates agent actions against executable financial rules and distinguishes attempted actions from enforced outcomes, while Fiducia-bench (https://arxiv.org/abs/2608.16055 ) uses machine-checkable policies and execution traces to study compliant behavior in financial workflows. Could the authors clarify more explicitly what evaluation capability this benchmark adds beyond these settings? I see the broader coverage of asset-management roles and business scenarios as a potential distinction, but would like to understand whether there is also a methodological difference.

  2. The paper shows that the choice of agent framework can substantially affect safety. Why were Hermes and OpenClaw chosen, and how representative are they of financial agent deployments? It would help to clarify how much the findings depend on these particular frameworks.

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

This paper constructs an asset-management agent-safety benchmark with three deployment settings, thirteen business subscenarios, and 877 cases. Cases specify identity, authority, and business state, and separately measure violations and lawful task completion. Experiments compare models, agent implementations, and added defenses; continuation and authorization-state studies show that the same runtime can have opposite safety effects across models, and initially safe behavior can fail under further interaction.

Critical strengths
  1. The benchmark checks both unauthorized behavior and lawful task completion, avoiding treating blanket refusal as safe completion. Execution-event scoring also detects cases where the final response appears appropriate but tool actions have already crossed a boundary, making the evaluation closer to business risk than final-text checks alone.

  2. Configuration comparisons and continuation experiments provide concrete, useful findings. In Table 4, the same agent implementation reduces one model's violation rate from 24.6% to 12.8% but raises another's from 7.9% to 19.4%; Table 7 shows that some initially safe runs violate under continuation, demonstrating the importance of model-runtime combinations and the interaction process.

Critical weaknesses
  1. Safe completion in the dynamic-state experiment is narrower than the main benchmark's compliance requirements. Section 4 includes required checks and evidential obligations in violations, while Table 8 and Section 6 state that the controlled analysis scores prohibited events only; with changed records and a neutral follow-up, Qwen3.8-Max under Minimal in S3.2 has 46.7% safe completion but only 33.3% record re-checking. The completion rate therefore does not directly establish that the agent reverified the current limit or safely adapted while satisfying procedural obligations, limiting the diagnostic depth of the state-change experiment.

  2. The scope of the tested defenses limits the applicability of those results. Table 6 groups generic attacks with overt malicious payloads separately from ordinary-looking financial requests seeking unauthorized outcomes, while the tested interventions focus on content detection, session-start audits, and tool restrictions without checking business state at commitment. The results mainly characterize coverage differences of these existing integrations across attack forms and do not directly establish the effectiveness of business-state-aware protection.

Questions for authors

In the S3.2 changed-record arm of Table 8, could you report safe completion conditioned on re-checking the deciding record, and identify which cases require a refresh? This would clarify how the controlled study relates to the procedural obligations in the main benchmark.

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

The paper designed an executable safety benchmark for financial asset-management agents, with 877 cases across customer service, internal assistance and autonomous trading settings. Across 144 model–agent–defense configurations, it shows that the agent harness can raise or lower safety depending on the model, and that added defenses barely help on financial attacks.

Critical strengths

The checkers score the agent from its actual tool calls and the final business state, and they measure violation and safe completion separately, so simply refusing a lawful task does not count as safe. The evaluation is across 9 models, 3 agent harnesses and 5 defenses with 3 seeds each, which is quite comprehensive. It shows that the same harness can make one model safer and another model less safe.

Critical weaknesses
  1. Most violations in the benchmark, like sending to an unauthorized address or ordering over a limit, could be blocked by the tool itself checking the business state, as real financial systems do. The paper’s tools do not enforce these rules, and the defenses tested are prompt-injection detectors, so it is not surprising they barely help. It is unclear to me whether this should be treated as an agent safety problem rather than a tool design problem.
  2. The checker’s violation precision and recall are about 86–88%, so it is wrong in roughly one out of eight cases, and the audit covers only two models. Many of the reported differences are only a few percentage points, such as the 1–4 point defense effects, so it is unclear whether these differences are larger than the checker’s own error.
Questions for authors
  1. How do the results change if the business tools enforce the permission and limit rules themselves? Which violations would still remain after that?
  2. In the record-change study, the record changes without any notice to the agent. Why should the agent be expected to look up the record again, instead of the system notifying it of the change?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.