playbook-case-study-page
Pass
Audited by Gen Agent Trust Hub on Aug 23, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites (
page_text). It mitigates risks through a robust verification framework: - Ingestion points: External prospect websites are scraped via the
scrape-websitetool. - Boundary markers: The prompt instructions explicitly command the model to ignore navigation, cookie banners, and irrelevant content while enforcing a strict JSON schema.
- Capability inventory: The skill is limited to data retrieval and string formatting; it lacks the ability to write to the file system, execute shell commands, or perform unauthorized network operations based on the scraped content.
- Sanitization: Five distinct "verbatim gates" implemented in JavaScript and Python ensure that the extracted customer names and evidence quotes exist exactly as found in the raw source text before any data is processed for shipment.
- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill performs HTTP probes and GET requests to external domains provided as input. These operations are limited to identifying public case-study pages (e.g.,
/customers,/testimonials) and do not involve the access or exfiltration of sensitive local files, credentials, or environment variables. - [COMMAND_EXECUTION]: The skill contains static JavaScript and Python code blocks intended for execution within the Clay platform's logic nodes. These scripts perform safe operations such as domain normalization, URL path construction, and data validation. They do not utilize dangerous functions like
eval()orexec()on untrusted external data.
Audit Metadata