The Agent Skills Directory

[SAFE]: The skill implements a robust verification process with explicit security guardrails. It instructs the agent to treat all external data as evidence rather than executable instructions, effectively mitigating potential injection attacks from the data being verified.- [DATA_EXFILTRATION]: Instructions specifically mandate the redaction of secrets, personal data, and access-bearing URLs, ensuring that sensitive information discovered during verification is not exposed.- [REMOTE_CODE_EXECUTION]: While the skill involves running benchmarks and application tests, it requires the use of isolated environments and clean disposable states. Specific sub-methods like MCP dependency review explicitly forbid the execution of untrusted commands or package installation during analysis.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as issue text, reports, and tool outputs (ingestion points in SKILL.md). It employs explicit boundary markers by instructing the agent to treat this data as evidence rather than instructions. While the skill's capability inventory includes running benchmarks and application reviews, it implements sanitization by requiring the redaction of sensitive data before reporting.- [PRIVILEGE_ESCALATION]: The skill maintains a read-only posture, directing the agent not to edit, repair, or commit changes to the candidate being verified, thereby preventing unauthorized modifications to the system.

kata-verify