The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from sources such as website scrapes and job listings (handled by scripts/analyze_signals_v2.py and scripts/analyze_signals.py). This data is analyzed to extract buyer language and signals. There is a risk that malicious instructions embedded in these external sources could influence subsequent agent actions if the extracted snippets are used in prompts without robust sanitization or delimiters.\n
Ingestion points: scripts/analyze_signals_v2.py, scripts/analyze_signals.py.\n
Boundary markers: The skill instructions mention "mining discovery documents without labels" but the scripts do not implement explicit prompt boundary markers for extracted text.\n
Capability inventory: scripts/find_contacts.py contains subprocess.run calls to external tools.\n
Sanitization: Content is normalized using unicodedata.normalize, but there is no explicit sanitization for potential prompt injection characters or sequences in the extracted evidence quotes.\n- [EXTERNAL_DOWNLOADS]: The SKILL.md provides instructions for installing the deepline CLI utility using npm install -g from a vendor-specific registry (https://code.deepline.com/api/v2/npm/). This involves downloading and executing installation scripts from a remote source.\n- [COMMAND_EXECUTION]: The scripts/find_contacts.py script executes the deepline command-line tool via subprocess.run. It builds commands using parameters derived from CSV input files. While it avoids shell execution and uses list-based argument passing, it relies on the integrity of the external CLI tool and the input data provided to it.

niche-signal-discovery