ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents
ToolHazard is meant to mass-produce hostile test environments for tool-using agents.
The paper says LLM agents can be compromised by indirect prompt injections hidden in the states of external tools. Its framework generates executable, stateful environments, finds injection points, creates payloads, and builds long-horizon tasks around them. The authors use it to create ToolHazard-Bench, where experiments show major agent vulnerabilities and sensitivity to when and where injections appear. They also report that ToolHazard-generated alignment data improves security on ToolHazard-Bench and AgentDojo without hurting benign task performance. HF Daily Papers' note
The paper says LLM agents can be compromised by indirect prompt injections hidden in the states of external tools. Its framework generates executable, stateful environments, finds injection points, creates payloads, and builds long-horizon tasks around them. The authors use it to create ToolHazard-Bench, where experiments show major agent vulnerabilities and sensitivity to when and where injections appear. They also report that ToolHazard-generated alignment data improves security on ToolHazard-Bench and AgentDojo without hurting benign task performance. HF Daily Papers' note
score 5