Show HN: AgenticBench, 9 security tests run on 25 AI coding agents

Category: security

Tags: ai-agents, security-benchmark, prompt-injection, sandboxing, evaluation

Score: 4.7/10 (Innovation: 5, Technical: 4, Documentation: 3, Utility: 5)

AgenticBench is a security-focused benchmark that runs nine prompt-injection and boundary-escape tests against 25 AI coding agents/harnesses, checking whether secrets leak or project content can override operator controls. It targets a real and under-tested gap (agent sandboxing and context isolation), but the README is essentially a leaderboard stub with test descriptions and offers full methodology only under NDA, limiting reproducibility and transparency.

Target audience: security engineers, AI agent developers, backend devs, devops

Repository: https://agenticbench.org/

View on Hacker News