skills/mukul975/anthropic-cybersecurity-skills/analyzing-browser-forensics-with-hindsight/Gen Agent Trust Hub
analyzing-browser-forensics-with-hindsight
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted browser activity records which can expose the agent to indirect prompt injection if malicious instructions are embedded within webpage titles, URLs, or metadata.
- Ingestion points: Local browser database files such as
History,Cookies,Web Data, and extensionmanifest.jsonfiles are parsed inscripts/agent.pyandscripts/process.py. - Boundary markers: Absent. There are no delimiters or defensive prompts instructing the agent to ignore any natural language instructions contained within the parsed database entries.
- Capability inventory: Limited to local file-write operations. The scripts write report files using
json.dumpinscripts/agent.pyandscripts/process.py. No network calls, subprocess execution, or dynamic code evaluations are present. - Sanitization: Absent. Webpage titles, URLs, and cookie attributes are extracted and handled directly as strings without sanitization or escaping, though autofill entries in
scripts/agent.pyare truncated to 50 characters.
Audit Metadata