# ClawSafety: "Safe" LLMs, Unsafe Agents

arXiv: https://arxiv.org/abs/2604.01438 · HTML: https://mlsanity.com/paper/2604.01438/

## Key facts

- Cited by 17 papers in the map (data through 14 Oct 2026).
- Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use.
- Builds on 28 of its 47 references that are in the map.
- First posted on arXiv 1 Apr 2026.

## Placed in

- LLM agents and tool use / [Agent threats, guardrails and governance](https://mlsanity.com/agents/agent-threats-guardrails-and-governance/)
- LLM safety and security / [Prompt injection and agent security](https://mlsanity.com/safety/prompt-injection-and-agent-security/)

## Builds on (most cited first, 28 of 28)

- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection](https://mlsanity.com/paper/2302.12173/) (2023)
- [HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face](https://mlsanity.com/paper/2303.17580/) (2023)
- [InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents](https://mlsanity.com/paper/2403.02691/) (2024)
- [AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents](https://mlsanity.com/paper/2410.09024/) (2024)
- [Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents](https://mlsanity.com/paper/2410.02644/) (2024)
- [Agent S: An Open Agentic Framework that Uses Computers Like a Human](https://mlsanity.com/paper/2410.08164/) (2024)
- [Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs](https://mlsanity.com/paper/2502.17424/) (2025)
- [Agent S2: A Compositional Generalist-Specialist Framework for Computer Use Agents](https://mlsanity.com/paper/2504.00906/) (2025)
- [WASP: Benchmarking Web Agent Security Against Prompt Injection Attacks](https://mlsanity.com/paper/2504.18575/) (2025)
- [SafeAgentBench: A Benchmark for Safe Task Planning of Embodied LLM Agents](https://mlsanity.com/paper/2412.13178/) (2024)
- [Introducing v0.5 of the AI Safety Benchmark from MLCommons](https://mlsanity.com/paper/2404.12241/) (2024)
- [AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?](https://mlsanity.com/paper/2602.03117/) (2026)
- [From Prompt Injections to Protocol Exploits: Threats in LLM-Powered AI Agents Workflows](https://mlsanity.com/paper/2506.23260/) (2025)
- [WebInject: Prompt Injection Attack to Web Agents](https://mlsanity.com/paper/2505.11717/) (2025)
- [textitAgents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks](https://mlsanity.com/paper/2504.00218/) (2025)
- [Agents of Chaos](https://mlsanity.com/paper/2602.20021/) (2026)
- [Prompt Injection Attacks on Agentic Coding Assistants: A Systematic Analysis of Vulnerabilities in Skills, Tools, and Protocol Ecosystems](https://mlsanity.com/paper/2601.17548/) (2026)
- [From Assistant to Double Agent: Formalizing and Benchmarking Attacks on OpenClaw for Personalized Local AI Agent](https://mlsanity.com/paper/2602.08412/) (2026)
- [WAInjectBench: Benchmarking Prompt Injection Detections for Web Agents](https://mlsanity.com/paper/2510.01354/) (2025)
- [Learning Personalized Agents from Human Feedback](https://mlsanity.com/paper/2602.16173/) (2026)
- [WebSentinel: Detecting and Localizing Prompt Injection Attacks for Web Agents](https://mlsanity.com/paper/2602.03792/) (2026)
- [Agents in the Wild: Safety, Society, and the Illusion of Sociality on Moltbook](https://mlsanity.com/paper/2602.13284/) (2026)
- [Mitigating Indirect Prompt Injection via Instruction-Following Intent Analysis](https://mlsanity.com/paper/2512.00966/) (2025)
- [Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety](https://mlsanity.com/paper/2603.10044/) (2026)
- [Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning](https://mlsanity.com/paper/2603.20198/) (2026)
- [Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning](https://mlsanity.com/paper/2602.13562/) (2026)
- [Safety Training May Persist Through Helpfulness Optimization in LLM Agents](https://mlsanity.com/paper/2603.02229/) (2026)

## Built on by (most cited first, 17 of 17)

- [AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security](https://mlsanity.com/paper/2605.29801/) (2026)
- [Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation](https://mlsanity.com/paper/2606.10749/) (2026)
- [Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries](https://mlsanity.com/paper/2607.10113/) (2026)
- [LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS Environments](https://mlsanity.com/paper/2605.10779/) (2026)
- [Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens](https://mlsanity.com/paper/2606.30755/) (2026)
- [LivePI: More Realistic Benchmarking of Agents Against Indirect Prompt Injection](https://mlsanity.com/paper/2605.17986/) (2026)
- [TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning](https://mlsanity.com/paper/2605.10038/) (2026)
- [Agent-ValueBench: A Comprehensive Benchmark for Evaluating Agent Values](https://mlsanity.com/paper/2605.10365/) (2026)
- [ASEval: Automated Trajectory-Level Security Testing for Autonomous Agents](https://mlsanity.com/paper/2605.22321/) (2026)
- [HarnessRisk: A Lifecycle-Oriented Benchmark for Agent Harness Safety](https://mlsanity.com/paper/2608.17597/) (2026)
- [Constraining Host-Level Abuse in Self-Hosted Computer-Use Agents via TEE-Backed Isolation](https://mlsanity.com/paper/2605.06393/) (2026)
- [When Routine Chats Turn Toxic: Unintended Long-Term State Poisoning in Personalized Agents](https://mlsanity.com/paper/2605.06731/) (2026)
- [SciHazard: A Benchmark for Measuring Scientific Safety Risks with Decomposed Harm Scoring](https://mlsanity.com/paper/2607.18665/) (2026)
- [Terminal Agents: A Survey of AI Agents in Command-Line Environments](https://mlsanity.com/paper/2608.20485/) (2026)
- [OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution](https://mlsanity.com/paper/2608.00677/) (2026)
- [Deny Without Disabling: Authorization-Paired Evaluation and Control for Multi-Agent Systems](https://mlsanity.com/paper/2610.00371/) (2026)
- [Lumen-bench: Cross-Lingual Safety Gaps and the Say-Do Problem in Tool-Calling Large Language Models](https://mlsanity.com/paper/2610.15134/) (2026)

## Abstract (arXiv)

Personal AI agents like OpenClaw run with elevated privileges on users' local machines, where a single successful prompt injection can leak credentials, redirect financial transactions, or destroy files. This threat goes well beyond conventional text-level jailbreaks, yet existing safety evaluations fall short: most test models in isolated chat settings, rely on synthetic environments, and do not account for how the agent framework itself shapes safety outcomes. We introduce CLAWSAFETY, a benchmark of 120 adversarial test scenarios organized along three dimensions (harm domain, attack vector, and harmful action type) and grounded in realistic, high-privilege professional workspaces spanning software engineering, finance, healthcare, law, and DevOps. Each test case embeds adversarial content in one of three channels the agent encounters during normal work: workspace skill files, emails from trusted senders, and web pages. We evaluate five frontier LLMs as agent backbones, running 2,520 sandboxed trials across all configurations. Attack success rates (ASR) range from 40% to 75% across models and vary sharply by injection vector, with skill instructions (highest trust) consistently more dangerous than email or web content. Action-trace analysis reveals that the strongest model maintains hard boundaries against credential forwarding and destructive actions, while weaker models permit both. Cross-scaffold experiments on three agent frameworks further demonstrate that safety is not determined by the backbone model alone but depends on the full deployment stack, calling for safety evaluation that treats model and framework as joint variables. Code and data will be available at: https://weibowen555.github.io/ClawSafety/.