Researchers released VulnGym, a benchmark built to test whether AI coding agents can actually find security vulnerabilities across a full code repository rather than in an isolated, pre-selected snippet. Most existing vulnerability-detection benchmarks hand a model a function or file already known to contain a bug and ask it to classify or describe the flaw, a much easier task than what a real security review requires, which is autonomously exploring an entire unfamiliar codebase to figure out where a vulnerability might even be. VulnGym instead pairs 184 real GitHub security advisories with the actual vulnerable versions of their repositories, covering 408 distinct vulnerability entries across 23 projects, and annotates each one with the specific line-level entry point, the critical operation that produces the vulnerability, and the full trace connecting them. That fine-grained annotation lets the benchmark score an agent on more than a binary found-it-or-not basis: it can separately evaluate whether an agent correctly localizes the vulnerable code, correctly identifies the operation that makes it exploitable, and correctly constructs a supporting trace of evidence, which is closer to what a human security reviewer would need to produce to justify a finding. The paper's headline result is a sobering one for anyone hoping to lean on coding agents for security review today: current agents remain limited at the full end-to-end task and are particularly weak at building accurate supporting evidence, even when they correctly flag that something is wrong. For teams evaluating whether an agentic tool like a security-scanning plugin is ready to replace or supplement manual code review, VulnGym is a useful reality check and a benchmark worth watching as agents improve, since it is explicitly designed to measure the repository-scale skill that actually matters in practice rather than a simplified proxy for it.