call-notes
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill represents a vulnerability surface for indirect prompt injection because it ingests untrusted data from sales transcripts, prospect websites, and LinkedIn profiles.
- Ingestion points: Processes raw sales call transcripts (SKILL.md), fetches prospect websites for context (SKILL.md), and reads LinkedIn profile data (SKILL.md).
- Boundary markers: The skill includes an explicit reference to
references/agent-security.md, which contains instructions to treat all fetched and pasted content as data, never as instructions, and to quote findings rather than summarizing them to prevent instruction propagation. - Capability inventory: The agent has file-writing capabilities (creating and updating
.agents/product-context.md) and browser tool access (Playwright). - Sanitization: The instructions in
references/agent-security.mdmandate HTML-escaping for interpolated values in emitted markup and prohibit placing untrusted content in sensitive contexts like<script>blocks. - [COMMAND_EXECUTION]: The skill uses the Playwright browser tool to access external websites, specifically LinkedIn, to verify stakeholder information. While this is a primary function of the skill, it involves automated interaction with external web environments based on potentially untrusted input from call transcripts.
- [DATA_EXFILTRATION]: There is a potential risk of data exfiltration if the agent is coerced via indirect prompt injection to use the browser tool to transmit session data or sensitive context to an attacker-controlled endpoint. The skill explicitly forbids transmitting passwords or credentials in
SKILL.md("never ask for, store, echo, or transmit a login or password for LinkedIn or any site") andreferences/agent-security.md("Redact tokens out of anything you echo back"), which acts as a mitigation strategy.
Audit Metadata