The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided system descriptions and failure modes to populate hazard logs. This ingestion of untrusted data represents a potential attack surface. However, the skill explicitly includes guardrails (Section 1.2 'Data Poisoning in STPA') requiring the agent to cross-reference LLM-generated actions against a static, deterministic safety rule engine to prevent poisoning attacks that might seek to hide critical failures.
[SAFE]: The skill serves as a template for safety engineering and does not include any executable code, shell commands, or network operations. It adheres to best practices by instructing the agent to operate within segmented environments if interfacing with cyber-physical systems and mandates cryptographic sign-offs for downgrading critical hazard severities.

hazard-analysis