Your AI agent's exit interview now has 214 questions — and it's failing most of them

agent-egress-bench is an open-source benchmark corpus of 214 test cases evaluating AI agent egress security, covering SSRF bypass, encoding evasion, DLP scanning gaps, denial-of-wallet, and MCP chain abuse. The project recently overhauled its applicability gating to test on observable surfaces rather than self-reported tool capabilities, closing loopholes where agents could dodge harder test variants by under-claiming what they inspect. Corpus is at v2.4 with scoring v2.5.