The Cold-Start Safety Gap in LLM Agents

• URL: http://arxiv.org/abs/2606.07867v1
• Date: Fri, 05 Jun 2026 21:58:19 GMT
• Title: The Cold-Start Safety Gap in LLM Agents
• Abstract summary: We study whether tool-calling agents are equally safe throughout a conversation.<n>We find they are most vulnerable at the start of a session and become substantially safer after a few regular agentic tasks.<n>We recommend having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests.
• Score: 25.95386115331387
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Are tool-calling LLM agents equally safe throughout a conversation? We discover they are not: agents are most vulnerable at the very start of a session and become substantially safer after a few regular agentic tasks -- a phenomenon we term the cold-start safety gap. To study this systematically, we introduce Safety Over Depth for Agents (SODA), a benchmark that controls how many regular agentic tasks the agent completes before encountering a safety threat, supporting up to 20 preceding tasks. Evaluating 7 models from 4 families, safety improves by 9--52% as the number of preceding regular agentic tasks increases from zero to twenty. Representation analysis confirms that model hidden states gradually shift toward a safety-aligned region as more preceding tasks are present. By systematically studying which part of the preceding conversation matters most, we find that the regular agentic tasks themselves are the primary driver of safety, while the agent's own prior responses have less effect on safety but are essential for preserving later utility. This conclusion is further supported by evaluation on open-source safety benchmarks (AgentHarm, Agent Safety Bench) and utility benchmarks (BFCL, API-Bank), confirming that warming up the agent with regular agentic tasks before deployment makes it safer and preserves full capability. Based on these findings, we recommend a simple deployment strategy: having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests mitigates the cold-start safety gap. Our code is available at https://github.com/Trustworthy-ML-Lab/Agent-Cold-Start-Safety-Gap

Related papers

• AgentLens: Interpretable Safety Steering via Mechanistic Subspaces for Multi-Turn Coding Agent [53.82005364479556]
  Coding agents based on large language models (LLMs) demonstrate remarkable autonomous capabilities.<n>Existing safety mechanisms mainly rely on external guardrails.<n>We propose AgentLens, a white-box defense framework that performs runtime safety detection and representation-level mitigation.
  arXiv  Detail & Related papers  (2026-06-21T21:13:55Z)
• Why Does Agentic Safety Fail to Generalize Across Tasks? [3.5870827272767496]
  AI agents are increasingly deployed in multi-task settings, where the task to perform is specified at test time, and the agent must generalize to unseen tasks.<n> Empirical evidence suggests that even when the ability to execute generalizes to unseen tasks, the ability to do so safely frequently does not.<n>This paper provides theory and experiments indicating that failures of agentic safety to generalize across tasks are not merely due to limitations of training methods, but reflect an inherent property of safety itself.
  arXiv  Detail & Related papers  (2026-05-07T22:16:03Z)
• ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback [53.2744585868162]
  Monitoring step-level tool invocation behaviors in real time is critical for agent deployment.<n>We first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents.<n>We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning.
  arXiv  Detail & Related papers  (2026-01-15T07:54:32Z)
• Too Helpful to Be Safe: User-Mediated Attacks on Planning and Web-Use Agents [3.7549350220109274]
  We study user-mediated attacks, where benign users are tricked into relaying untrusted or attacker-controlled content to agents.<n>We conduct a systematic evaluation of 12 commercial agents in a sandboxed environment.<n>Our results show that agents are too helpful to be safe by default.
  arXiv  Detail & Related papers  (2026-01-14T03:29:13Z)
• OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety [58.201189860217724]
  We introduce OpenAgentSafety, a comprehensive framework for evaluating agent behavior across eight critical risk categories.<n>Unlike prior work, our framework evaluates agents that interact with real tools, including web browsers, code execution environments, file systems, bash shells, and messaging platforms.<n>It combines rule-based analysis with LLM-as-judge assessments to detect both overt and subtle unsafe behaviors.
  arXiv  Detail & Related papers  (2025-07-08T16:18:54Z)
• Agent-SafetyBench: Evaluating the Safety of LLM Agents [72.92604341646691]
  We introduce Agent-SafetyBench, a benchmark designed to evaluate the safety of large language models (LLMs)<n>Agent-SafetyBench encompasses 349 interaction environments and 2,000 test cases, evaluating 8 categories of safety risks and covering 10 common failure modes frequently encountered in unsafe interactions.<n>Our evaluation of 16 popular LLM agents reveals a concerning result: none of the agents achieves a safety score above 60%.
  arXiv  Detail & Related papers  (2024-12-19T02:35:15Z)
• SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents [58.65256663334316]
  We present SafeAgentBench -- the first benchmark for safety-aware task planning of embodied LLM agents in interactive simulation environments.<n>SafeAgentBench includes: (1) an executable, diverse, and high-quality dataset of 750 tasks, rigorously curated to cover 10 potential hazards and 3 task types; (2) SafeAgentEnv, a universal embodied environment with a low-level controller, supporting multi-agent execution with 17 high-level actions for 9 state-of-the-art baselines; and (3) reliable evaluation methods from both execution and semantic perspectives.
  arXiv  Detail & Related papers  (2024-12-17T18:55:58Z)
• GuardAgent: Safeguard LLM Agents by a Guard Agent via Knowledge-Enabled Reasoning [79.07152553060601]
  We propose GuardAgent, the first guardrail agent to protect target agents by dynamically checking whether their actions satisfy given safety guard requests.<n>Specifically, GuardAgent first analyzes the safety guard requests to generate a task plan, and then maps this plan into guardrail code for execution.<n>We show that GuardAgent effectively moderates the violation actions for different types of agents on two benchmarks with over 98% and 83% guardrail accuracies, respectively.
  arXiv  Detail & Related papers  (2024-06-13T14:49:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.