The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts using the python -m and python scripts/ patterns to automate data collection (scripts.audit.collect) and perform quality/compliance gating (scripts/quality_gates/four_us_score.py, scripts/quality_gates/banned_word_check.py).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data from sources such as public review platforms (Google, G2, Trustpilot), emails, DMs, and call transcripts, which may contain malicious instructions aimed at the agent.
Ingestion points: Data enters the system via kai-data.json (output from the collector script) and various external assets (reviews, emails, transcripts) processed in Phase 1.
Boundary markers: The skill employs [S] (sourced) and [C] (connective tissue) markers during Phase 4 fusion to distinguish between source data and generated text, and enforces verbatim quote rules, though these are primarily for accuracy rather than security sandboxing.
Capability inventory: The agent has the capability to execute shell commands (via Python scripts) and write files to the workspace/ directory.
Sanitization: The skill includes a 'Compliance Gate' (Phase 5) using quality scoring scripts and checks against FTC rules to validate the content, though these do not specifically target prompt injection patterns.
[DATA_EXFILTRATION]: The skill instructions direct the agent to 'search for email/ad/analytics config' and access sensitive internal sources including CRM exports, billing systems, and product database queries. When combined with the network-enabled collector script (scripts.audit.collect --url), this creates a potential path for data exposure or exfiltration if the agent is misled regarding the destination URL or data handling.

kai-proof-builder