LLM-enabled agents are gaining popularity but introduce new security vulnerabilities, particularly TOCTOU flaws. TOCTOU vulnerabilities involve state validation followed by modification before use, leading to exploitable issues. This research presents the first study specifically focusing on TOCTOU vulnerabilities within LLM agents. The authors introduce TOCTOU-Bench, a benchmark with 66 realistic user tasks designed to detect TOCTOU vulnerabilities. They adapted detection and mitigation techniques from systems security for application in this context. These countermeasures include prompt rewriting, state integrity monitoring, and tool-fusing strategies. Automated detection methods achieve up to 25% detection accuracy within agentic workflows. The mitigation strategies reduce vulnerable plan generation and the attack window significantly. Combining all three methods decreased the incidence of TOCTOU vulnerabilities in executed trajectories. The study emphasizes the unique challenges of securing agentic workflows against TOCTOU attacks. This work aims to open new avenues for research in AI safety and systems security.