CyCraft Unveils Key Agentic AI Security Risks and Evaluation Benchmark

As companies accelerate the adoption of agentic AI, systems are gaining the ability to make decisions autonomously and invoke external tools, creating new cybersecurity vulnerabilities. Traditional static security testing struggles to prevent malicious attacks during execution, driving an industry consensus that agent security is runtime security. Real-time tracking of an AI agent’s intermediate steps while it performs a task has become a top priority for protecting enterprises against prompt injection and tool misuse.

Taiwanese cybersecurity company CyCraft held its annual technology conference on July 3, 2026, where it jointly released the TraceSafe research paper with National Taiwan University. The paper was also presented at ICML 2026. The research identified 12 key risks associated with tool calls and introduced TraceSafe-Bench, an evaluation benchmark containing more than 1,000 execution cases to help companies address cybersecurity challenges.