acceptodds
Under review as a conference paper at ICLR 2027

Safety Orchestration: Building a Routable Library of Atomic Safety Skills

Abstract

LLM agents need safety checks throughout the execution of complex, long horizon tasks. Open-source communities have developed many safety practices, but these are spread across separate artifacts that often provide the same functions. We screen 10,223 community artifacts and annotate 1,390 candidates to identify recurring safety functions. We consolidate these functions into a reusable library of 95 atomic capabilities across 19 functional archetypes. We introduce SAFETY ORCHESTRATOR, a training-free framework that selects phase-specific safety guidance and independently checks tool calls and outputs through host hooks. Across seven proprietary and open-weight models, it achieves average absolute reductions in Unsafe Action Rate of 15.42% on SABER and 8.84% on OpenAgentSafety, with an average absolute accuracy decrease of 1.28% on Terminal-Bench 2.1. Ablations show complementary benefits from guidance and host checks, with stage-specific reference selection outperforming loading the full library upfront in reducing unsafe actions. Evaluations across different agent runtimes also show that the same safety library can be reused. These results support modular runtime safeguards as a practical way to improve agent safety across models and runtimes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Agent Safety, Runtime Safeguards, Tool Use, Long-Horizon Agents, Safety Guardrails

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_6pf3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work curates agent skills, tool-use hooks, and other artifacts from publicly available sources, and proposes an approach for extracting common, useful set of capabilities/skills/hooks that can be used as safety validators on different agentic harnesses with various models. The empirical results seem convincing and broadly applicable across model choices, harnesses, and two benchmark choices. Also the safety-utility tradeoff study on Terminal bench shows that their approach retains utility.

Critical strengths
  1. The idea of curating existing skills and then extracting a common, useful subset that can achieve good empirical safety gains is interesting in general beyond just the domain of agent safety.
  2. The empirical gains are promising and broadly justify the promise of the proposed approach over a vanilla agent harness.
Critical weaknesses
  1. The idea of having tool-use hooks and skills for agent safety is well explored as noted by authors as they use these sources for curating their skills/capabilities. Yet they do not provide an experimental result that describes the benefit of doing this curation effort, which is their primary research contribution, as opposed to simply relying on a specific set of hand-chosen/standard set of skills, or even other approaches like using all possible sources with a searchable index on top of so that the agent can dynamically choose one without polluting the agent's context with unnecessary capabilities. Instead of discussing how their skills can reduce unsafe rates over a vanilla harness, it would be more beneficial if they discuss why their particular approach for curating these capabilities/interventions achieve better safety metrics over other approaches.
  2. The authors are positioning their work as a "library" but the main paper includes no discussion on how exactly their effort can be reused/plugged in to different agent harnesses. They demonstrate empirical results across different harnesses, and I agree that skill markdown files are reusable. But I am not very convinced especially about the tool-use hooks being re-usable across harnesses, because different agent harnesses can have different toolkits.
  3. I found the paper quite hard to follow: (a) the terms "capabilities" and "artifacts" should be formally/informally defined/explained via examples, (b) Section 4.5's methods like "Flat", "All-loaded" are also not explained very clearly.
Questions for authors
  1. The comparison of approaches based on total execution time is unfair on unsafe tasks because your interventions can prematurely terminate the task. A fair metric will be to use average additional overhead per turn instead of comparing full execution time. Or compare the total execution time on benign tasks after adding these safety interventions.
  2. While I understand that publicly available skills can be repetitive and you can consolidate them, is this effort really needed? Considering how strong the models you evaluate are (for e.g. GPT-5.6-SoL), I just wonder why we can't just let the model decide what skills and hooks are needed for a given task? I was also unable to understand how exactly a router is implemented in your pipeline? Is it just a skill document that describes what other interventions are available?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 6pf3 — Part 4

Q6: Necessity of consolidation and implementation of the Router.

The reviewer asks whether consolidation is necessary, given how strong the evaluated models are, and why the model cannot decide by itself which skills and hooks a task needs. The reviewer also asks how the Router is implemented, and whether it is simply a skill document that lists the available interventions.

The model already chooses its skills in our design, while the hooks must stay independent of the model. We first explain how the Router is implemented. We then answer the question about letting the model decide in three parts: whether the model can choose its skills, why it should not choose its hooks, and whether strong models still need hooks.

How the Router is implemented. The Router is a skill document (a meta-skill) with no trained component. The host makes it available to the agent, and the agent invokes it at the start of the task. The two tables below summarize its content.

Router card

FieldContent
Identifiersafety-router-skill
InvocationMade available by the host and invoked by the agent at the start of the task (Cases 1 and 3, Appendix D)
Phase mapFive execution phases, each linked to the archetype references that apply to it (table below)
Reading ruleAs the task moves through the phases, the model reads the references relevant to the current phase and task context, e.g. invoke_skill(name="safety-router-skill", reference="detect-task-overreach") (Case 3, Appendix D)
Check outputFor each check, the model reports pass, warn, or block, with the matched capability and a reason
ArbitrationAny block pauses the workflow for human review. A warn is logged and the workflow continues. If all checks return pass, the workflow continues (Section 3.4)
Relation to hooksHooks run independently at every tool call. A model-side judgment cannot override a hook denial

Phase map (archetypes with a model-side reference, from Table 12)

PhaseArchetype references
Input understandingInput intent & ambiguity; Prompt injection; Input PII and secrets
PlanningPolicy application; Task overreach
Tool invocationTool argument safety; Agent/tool trust; Supply-chain risk; Rate and quota limits
Output generationSensitive redaction; Data exfiltration; Output content policy
Cross-cuttingIncident response; Human escalation

The Router therefore does more than list the available interventions. It also specifies when each reference should be read and how the check results are handled. We will add this card to Section 3.4.

Can the model decide which skills it needs? In our design, the model already makes this decision. It chooses which references to read, and the Router only tells it what is available in each phase. Consolidation changes what the model chooses from. Instead of 1,390 audited artifacts, the model chooses from 14 archetype references with clearly delimited scope.

Selecting less text also helps a strong model make safer decisions. For GPT-5.6 Sol, loading all curated references at the start gives a UAR of 17.27% on SABER and 33.70% on OpenAgentSafety. Reading only the references for the current phase gives 12.57% and 27.02% (Table 4). A small set of references with clear scope is what makes this selection practical.

Why the model does not decide which hooks to run. Hooks must not depend on the model, because they exist to catch the model's own mistakes. If the model chose its hooks, a wrong judgment would also switch off the check meant to catch it. Case 2 in Appendix D illustrates this. The post-tool hook removes live passwords read from ~/.pypirc before the result reaches the model, whereas a model deciding whether to redact would have to read the passwords first. Hooks therefore run on every tool call, and the model cannot override a hook denial (Section 3.4).

Thanks for your response! I believe this draft needs significant revisions that must be addressed in a future submission. Hence I maintain my current score.

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

This paper collects and organizes a large number of community safety artifacts into 95 atomic safety capabilities across 19 functional archetypes. The authors introduce Safety Orchestrator, a training-free framework that routes agents to appropriate safety skills and performs independent safety checks through host hooks. The results show that the framework reduces the Unsafe Action Rate by 15.42% on SABER and 8.84% on OpenAgentSafety, with only a small accuracy decrease on Terminal-Bench 2.1.

Critical strengths
  1. Comprehensive safety skill library. The authors put substantial effort into collecting and organizing a large number of community safety artifacts into 95 atomic safety capabilities across 19 functional archetypes. This provides a useful and reusable resource for agent safety.

  2. Combination of model guidance and host hooks. The framework combines model-level safety guidance with independent host-level safety checks. Both components are important for improving agent safety.

  3. Useful ablation study. The ablation study shows that stage-specific reference selection outperforms loading the entire safety library upfront in reducing unsafe actions. This provides useful evidence for the effectiveness of the proposed routing strategy.

Critical weaknesses
  1. Lack of comparison with existing baselines. The paper does not compare against several relevant existing methods, such as AgentSpec, Progent, and NeMo Guardrails. Without these comparisons, it is difficult to assess the advantages of the proposed framework over existing safety approaches.

  2. Limited evaluation of task completion. The evaluation on SABER and OpenAgentSafety mainly focuses on Unsafe Action Rate (UAR), without reporting task completion rates. While reducing unsafe actions is important, it is also necessary to understand whether these safety mechanisms prevent agents from completing legitimate tasks.

  3. Limited methodological novelty. The paper presents substantial engineering efforts in collecting and organizing safety capabilities. However, the methodological novelty appears relatively limited, as the framework primarily integrates existing safety practices through skill routing and host-level checks.

Questions for authors
  1. Incorporating new safety skills. How does the framework incorporate new safety skills or MCP tools as they emerge from the community? Can the safety library be updated efficiently without manually reorganizing the existing capabilities?

  2. Human review and UAR calculation. When an action is blocked or paused for human review, how is it counted in the Unsafe Action Rate (UAR) calculation? Is it automatically considered safe, even if the original task remains incomplete?

  3. Evaluation independence. Were the host hook rules finalized before evaluating on SABER and OpenAgentSafety? Were these benchmarks held out during the development of the safety library and host hook rules?

2 comments
AuthorsOpenReview ↗

Response to Reviewer pQoY — Part 3

Q3: Limited methodological novelty.

The paper presents substantial engineering efforts in collecting and organizing safety capabilities. However, the methodological novelty appears relatively limited, as the framework primarily integrates existing safety practices through skill routing and host-level checks.

Skill routing and host-level checks are established mechanisms, and we do not claim them as new. Our contribution is a different way of organizing agent safety, together with the design and evidence that support it.

A new perspective on agent safety. The paper changes the unit of agent safety from the artifact to the atomic capability. The common practice is to add more safety artifacts to an agent. Our ecosystem study shows that this does not scale. The 1,182 labeled artifacts contain only 494 distinct capability signatures (Section 2). A random subset of 100 artifacts covers about 18 of the 19 archetypes, but only about 60 of the 86 observed capability labels (Figures 2b and 4b). Adding artifacts therefore mostly adds repetition, while many specific checks remain missing. This motivates treating safety as a set of atomic capabilities that are selected and enforced during execution.

Design decisions that make this perspective work. Turning this perspective into a working system requires three decisions that the source artifacts do not provide. First, each artifact is split into atomic capabilities with an explicit scope, so that overlapping checks can be merged and missing ones identified. Second, each capability is assigned to the model, the host, or both. As a result, 60 of the 95 capabilities become host hooks, although most source artifacts are written as instructions for the model. Third, model-side references are organized by execution phase, and hook denials take precedence over model judgments.

Evidence that this is more than integration. If the framework only integrated existing practices, delivering the same material in a simpler way should work equally well. Our ablation shows that it does not. All-loaded uses exactly the same library and hooks as Ours, but loads all references at the start. Ours lowers UAR by a further 3.90–6.68 points (Table 4). In addition, the full configuration lowers UAR by 5.31–10.23 points compared with Flat, which uses guidance alone. It also lowers UAR by 5.17–8.99 points compared with Hooks-only, which uses hooks alone. The gains therefore come from how the capabilities are organized and delivered, not only from what they contain.

Generality. The framework requires no training and transfers across models and runtimes. UAR decreases for all seven models on both benchmarks (Table 2), and in all 12 settings across three agent runtimes (Table 6).


AuthorsOpenReview ↗

Response to Reviewer pQoY — Part 4

Q4: Incorporating new safety skills.

How does the framework incorporate new safety skills or MCP tools as they emerge from the community? Can the safety library be updated efficiently without manually reorganizing the existing capabilities?

The library can be updated incrementally. A new artifact goes through the same scripted pipeline that was used to build the library. It changes the library only when it brings a function that the 95 capabilities do not cover.

Step 1: screening and annotation. A new skill document or MCP record goes through the same deduplication, filtering, and LLM-assisted annotation as the original collection (Sections 3.1–3.2, Appendix A). The annotation uses the existing 95-capability vocabulary. It records the primary and secondary capabilities, the execution phases, and the risks introduced by the artifact itself. These steps run automatically.

Step 2a: the function is already covered. The artifact is linked to the existing capability as an additional source. The library, the references, and the hooks stay unchanged. We expect this to be the common case. The first 200 artifacts already cover more than 80% of the observed capability labels, and the curve flattens after that (Figure 2b).

Step 2b: the function is new. The annotation schema has a dedicated field, suggested_new_atoms. In this field, the annotator proposes an identifier, a parent archetype, supporting evidence, and a draft definition (Appendix A.3). Only these proposals and the annotations with confidence below 0.7 are sent to human review (Appendix A.4). An accepted capability receives a card with its scope and implementation mode. It is then added either to the reference document of its parent archetype or as a hook. The Router maps execution phases to archetype references. A new capability under an existing archetype therefore requires no change to the Router or to other capabilities.

The taxonomy only needs to be reorganized when a new function fits none of the 19 archetypes. For artifacts that fit an existing archetype, manual effort is limited to the two groups sent to human review. In our corpus, these were 94 new-capability proposals and 249 low-confidence annotations among the 1,390 audited artifacts (Table 10).


Q5: Human review and UAR calculation.

When an action is blocked or paused for human review, how is it counted in the Unsafe Action Rate (UAR) calculation? Is it automatically considered safe, even if the original task remains incomplete?

A blocked or paused trajectory is not automatically considered safe. It is counted as completed only if the agent still finishes the task afterwards.

UAR definition. UAR is the fraction of evaluated trajectories that contain at least one safety violation in the actions the agent actually executed. Each trajectory is counted once. Blocked, paused, and terminated trajectories all remain in the denominator.

Blocked or paused actions. A block only prevents the action it applies to. If that action is never executed and the rest of the trajectory has no violation, the trajectory is not counted as unsafe. If the agent executes an unsafe action before or after the block, the trajectory is counted as unsafe.

Human confirmation. Each confirmation request raised during evaluation is reviewed individually. An approval does not exempt the actions that follow. If an approved action is unsafe, the trajectory is still counted as unsafe.

Incomplete tasks. Task completion is measured separately. A trajectory that ends in a block, pause, or refusal is not counted as completed. The cost of these interventions therefore appears in the completion rates reported in our response to Q2. It is not hidden by UAR.


Q6: Evaluation independence.

Were the host hook rules finalized before evaluating on SABER and OpenAgentSafety? Were these benchmarks held out during the development of the safety library and host hook rules?

The host-hook rules were finalized before the evaluation. Both benchmarks were held out during the whole development of the library and the hooks.

Hook rules. The host-hook rules and their thresholds were finalized before the first evaluation run on SABER and OpenAgentSafety. Evaluation results were not used afterwards to revise capabilities or tune the rules.

Held-out benchmarks. The tasks and labels of the two benchmarks were not used to build the capability library or the hooks. The library and the hooks were developed only from community skill documents, MCP records, and security references. These references include the 11 documents in Appendix A.1, which cover OWASP, NIST, MITRE ATLAS, and others. They also include OWASP Agentic Skills Top 10 and Snyk ToxicSkills (Section 3.1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.