agent-redteam-lab

To test an agent for prompt injection, you run structured red-team scenarios that inject hostile document text to see if it can override higher-priority instructions and bypass boundary protection. The Skill then generates an evidence-based report.

Canary testing for agent safety checks for synthetic secret leakage, memory poisoning, and hidden prompt or policy disclosure. You need it during defensive red-team validation to ensure sensitive boundaries remain intact against hostile instructions.

Yes, you can validate tool boundaries for a local MCP-connected agent. The Skill verifies that sensitive reads, shell actions, network egress, and other high-impact tools remain properly gated against misuse.

The best way to perform red-team validation on a published agent API is to apply structured scenarios with strict scenario isolation and benign controls. This approach preserves tool, secret, memory, and network boundaries while producing severity-labeled reports.

An agent fails sensitive tool boundary checks during hostile document review when it improperly executes sensitive reads, shell actions, or network egress in response to injected text. Evidence-based severity labeling identifies these exact boundary protection failures.

Yes, agent red-team testing requires scenario isolation along with benign controls. This strict isolation preserves tool, secret, memory, and network boundaries, ensuring evidence-based severity labeling and accurate remediation guidance.