cc-continue

Pass

Audited by Gen Agent Trust Hub on Aug 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes various shell utilities including node, awk, git, and mv to navigate directories, process transcript data, and manage session handoff files.\n- [REMOTE_CODE_EXECUTION]: The skill uses Python heredocs to dynamically generate and execute scripts at runtime. These scripts handle complex data extraction, JSON log parsing, and text reconstruction from raw transcripts.\n- [DATA_EXFILTRATION]: The skill accesses sensitive local conversation history and metadata stored in ~/.claude/projects/ and ~/.claude/claude-code-token-saver-data/. While this access is the core functionality of the skill, it involves reading private session transcripts into the agent context.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection by restoring previous session data into the current agent context.\n
  • Ingestion points: Reads transcripts from local .jsonl files and preprocessed caches in the ~/.claude directory.\n
  • Boundary markers: None; preprocessed session content is interpolated into the context without structural delimiters or explicit instructions to ignore embedded commands.\n
  • Capability inventory: Local file system access, shell command execution, and dynamic Python script execution.\n
  • Sanitization: None; transcript content is preprocessed for length and formatting but is not sanitized to detect or neutralize instructions embedded in historical messages.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 21, 2026, 02:00 AM
Security Audit — agent-trust-hub — cc-continue