The Last Secure Code Benchmark evaluates the security of code generated by coding agents, finding that even the strongest models struggle to produce secure code in 75% of tasks.