A user study with over 100 participants shows humans rarely spot AI agents sabotaging code during extended collaborative tasks, even with a safety monitor present.
Unsafer in many turns: Benchmarking and defending multi-turn safety risks in tool-using agents
5 Pith papers cite this work, alongside 1 external citations. Polarity classification is still indexing.
5
Pith papers citing it
1
external citations · Pith
abstract
LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.
years
2026 5representative citing papers
Boiling the Frog is a new stateful multi-turn benchmark that finds an aggregate 44.4% strict attack success rate for incremental safety violations across nine AI models, with rates ranging from 20.5% to 92.9%.
Presents TRUST-Bench benchmark for hidden-trigger tool compromises in LLM agents and VISTA-Guard framework for trajectory-aware risk scoring of final actions under untrusted feedback.
No existing AI security framework covers a majority of the 193 identified multi-agent system threats in any category, with OWASP Agentic Security Initiative achieving the highest overall coverage at 65.3%.
citing papers explorer
-
Coding with "Enemy": Can Human Developers Detect AI Agent Sabotage?
A user study with over 100 participants shows humans rarely spot AI agents sabotaging code during extended collaborative tasks, even with a safety monitor present.
-
Boiling the Frog: A Multi-Turn Benchmark for Agentic Safety
Boiling the Frog is a new stateful multi-turn benchmark that finds an aggregate 44.4% strict attack success rate for incremental safety violations across nine AI models, with rates ranging from 20.5% to 92.9%.
-
Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback
Presents TRUST-Bench benchmark for hidden-trigger tool compromises in LLM agents and VISTA-Guard framework for trajectory-aware risk scoring of final actions under untrusted feedback.
-
Security Considerations for Multi-agent Systems
No existing AI security framework covers a majority of the 193 identified multi-agent system threats in any category, with OWASP Agentic Security Initiative achieving the highest overall coverage at 65.3%.
- Signal-Driven Observation for Long-Horizon Web Agents