usage-view

Warn

Audited by Gen Agent Trust Hub on Oct 7, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DATA_EXPOSURE]: The skill accesses highly sensitive conversation transcripts from the local filesystem to generate reports.
  • It reads JSONL files from ~/.claude/projects/, which contain the full history of a user's interactions with the AI, potentially including code, secrets, and private discussions.
  • [DYNAMIC_EXECUTION]: The skill uses runtime instruction generation and self-modification of configuration files.
  • The agent is instructed to run a script that generates an agentPromptFile, then read and follow the instructions in that file exactly.
  • If a model is missing from its local database, the agent fetches pricing data from the web and is instructed to edit the skill's own model-pricing.json file to add the new entries.
  • [EXTERNAL_DOWNLOADS]: The skill fetches content from external sources to update its pricing data and render charts.
  • It fetches model pricing from Anthropic's documentation site at https://platform.claude.com/docs/en/about-claude/pricing#model-pricing.
  • The HTML dashboard template loads the Chart.js library from a public CDN (cdn.jsdelivr.net).
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from user conversation transcripts and passes it to an LLM sub-agent for analysis, creating an attack surface for indirect injection.
  • Ingestion points: Transcripts located in ~/.claude/projects/.
  • Boundary markers: The prompt uses behavioral constraints (e.g., "NEVER invent or modify numbers") but does not provide explicit delimiters to separate transcript data from instructions.
  • Capability inventory: Execution of local scripts, file system read/write access, network operations via WebFetch, and opening a browser.
  • Sanitization: Transcript content is not sanitized for potential malicious instructions before being analyzed by the sub-agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 7, 2026, 05:40 PM