The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes job listings, titles, and descriptions from a third-party source (Jobdanmark.dk), which is a common vector for indirect prompt injection. To mitigate this risk, the skill includes comprehensive sanitization logic:
Ingestion points: The skill fetches data via the Jobdanmark API in cli/src/commands/search.ts and scrapes job detail pages in cli/src/commands/detail.ts.
Sanitization: The stripInvisible function in cli/src/helpers.ts recursively cleanses all output by removing C0/C1 control characters, bidirectional marks, zero-width spaces, and Unicode tag characters (U+E0000-E007F) which can be used to hide instructions.
HTML Cleaning: The visibleHtml function in cli/src/helpers.ts strips non-rendering content such as HTML comments, script/style blocks, and elements marked as hidden via attributes or inline CSS before the agent sees the text.
Capabilities: The skill's capabilities are limited to network requests to the target domain and local processing of job data.
[OBFUSCATION]: Static analysis flagged the presence of steganographic Unicode tag characters in cli/src/helpers.ts and cli/tests/hidden-text.test.ts. This was determined to be a false positive. The characters are used in a regular expression in helpers.ts to remove them from external data, and in hidden-text.test.ts as a test fixture to verify that the security sanitization correctly identifies and strips such content.

jobdanmark-search