agent-egress-bench

Adopt it with reservations if you need a starting point for AI agent egress-security testing, not an out-of-the-box production defense. The available evidence is limited to one GitHub Search Lead, whose repository snippet describes it as a “tool-neutral attack corpus for AI agent egress security.” That supports the project’s stated direction, but does not establish coverage, detection accuracy, or operational effectiveness. The most accurate analogy is a security regression corpus or benchmark suite—not a security gateway, runtime firewall, or hosted scanning service.

Its intended practical role is to provide reusable attack inputs and scenarios for evaluating controls around data exfiltration, prompt injection, SSRF, secret detection, and MCP-related risks. Being tool-neutral suggests that the test material is not tied to one agent framework, model vendor, or protection product. However, the supplied evidence does not show the corpus size, file formats, execution commands, scoring methodology, baseline results, or reproducibility guidance.