Researchers have introduced ToolHazard, a scalable framework designed to synthesize adversarial environments for evaluating and aligning LLM-based agents. Recognizing that existing security research is often limited by manual environment engineering and predefined injection points, ToolHazard utilizes a multi-agent system—comprising an Environment Simulator, an Attacker Agent, and a User Simulator—to automate the creation of executable, stateful environments. This system identifies viable injection points within environmental states and generates domain-specific payloads to construct complex, long-horizon tasks. To facilitate testing, the authors developed ToolHazard-Bench, a benchmark designed to stress-test agent security across diverse workflows. Empirical evaluations demonstrate that agent vulnerabilities are significant and that the efficacy of an attack is highly dependent on the timing and placement of the injection. Furthermore, the study shows that alignment data synthesized via the ToolHazard framework enhances agent security against attacks on both ToolHazard-Bench and AgentDojo. Crucially, this security enhancement is achieved without compromising the agents' performance on benign, non-adversarial tasks, suggesting a viable path for scaling robust security evaluations in agentic systems.