The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted job data from LinkedIn's public guest endpoints. To protect the agent from hidden instructions, it implements comprehensive sanitization in cli/src/helpers.ts. The visibleHtml function strips comments, scripts, and elements marked as hidden from the source HTML. The stripInvisible function removes zero-width characters, bidirectional overrides, and Unicode tag characters (U+E0000–U+E007F) which can be used to hide malicious instructions in third-party text.
Ingestion points: Untrusted job listing HTML is retrieved in cli/src/helpers.ts via the htmlFetch function.
Boundary markers: Sanitization functions are applied at the final output boundary in runSearch and runDetail before data is returned to the agent.
Capability inventory: The skill is restricted to network access for LinkedIn's guest API and standard output; it lacks file system write access or persistence mechanisms.
Sanitization: Highly robust sanitization is implemented as a core feature of the skill logic.
[EXTERNAL_DOWNLOADS]: The skill fetches data from official LinkedIn guest endpoints. It includes strict host validation in cli/src/commands/detail.ts to prevent connecting to look-alike domains or using userinfo tricks for phishing. Automated scan alerts for phishing URLs (e.g., linkedin.com.evil.io) are successfully mitigated as the skill's normalizeId function explicitly identifies and rejects these patterns during validation.
[OBFUSCATION]: Static analysis flags for Unicode steganography in cli/src/helpers.ts and cli/tests/hidden-text.test.ts are verified to be defensive measures. The skill defines these characters in a regular expression specifically to strip them from external input, and the test files contain them to verify the effectiveness of the sanitization logic.

linkedin-search