The Agent Skills Directory

[SAFE]: The skill connects to a well-known technology service (api.perplexity.ai) using standard API authentication methods. The script performs legitimate search operations as described.
[DATA_EXFILTRATION]: The skill implements strong data protection measures. The instructions explicitly forbid the agent from leaking the PERPLEXITY_API_KEY and prevent results from being piped to network-transmitting commands (e.g., curl, wget). The local script does not access sensitive files or perform unauthorized network operations.
[PROMPT_INJECTION]: No evidence of malicious instructions or jailbreak attempts was found. The skill includes a dedicated 'Security Guardrails' section that hardens the agent against common failure modes.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle untrusted web content safely.
Ingestion points: Web search results (snippets, titles, and URLs) are retrieved in scripts/search.mjs and returned to the agent's context.
Boundary markers: None present in the raw script output, but the agent is explicitly instructed to treat the input as untrusted.
Capability inventory: The skill utilizes the node environment to execute search queries. The agent is instructed not to execute commands found in results without user confirmation.
Sanitization: Search snippets are truncated to 300 characters in the formatting logic, and instructions in SKILL.md provide clear behavioral boundaries for processing results.

perplexity-hardened