invarislabs/invaris-agentsec is a Python testing framework that checks whether security policies hold throughout AI agent workflows. The latest release on the GitHub repository is v0.9.0, licensed under Apache-2.0. It covers adversarial scenarios including prompt injection, tool misuse, memory poisoning, and MCP scanning.
The tool tracks whether an attack payload actually reaches the agent. Runs in which it does not reach the agent are treated as inconclusive rather than passing, and only tool calls made after content delivery count when judging success. Checked items include unauthorized tool calls, poisoning through tool outputs and MCP servers, excessive tool calls, tokens, and cost, and infinite loops. It also covers dangerous call combinations, identity and session confusion, subagent privilege abuse, and cases where an agent denies an action or fails to report something it did. It also provides regression tests and reports for CI.
The README explains that scenarios reference commands, policy files, and multi-agent, email, and browser pack documents, with the checklist and documentation corresponding item by item. However, the figures concerning risk scenarios and benchmarks are the README's own claims. So while the check design and scope are specific, the security effect cannot be independently confirmed from this record alone.