context-mode

Pass

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill's primary function is to execute arbitrary code across multiple languages (JavaScript, Python, Shell, etc.) using tools like ctx_execute. The documentation clearly states that these processes inherit the user's filesystem and network authority and are not sandboxed. It explicitly instructs the agent to obtain user approval before any execution.
  • [REMOTE_CODE_EXECUTION]: The instructions include commands to install and run the context-mode package from the npm registry using npm install -g and npx -y. These commands use specific version pins (e.g., 1.0.169) to ensure the integrity of the downloaded code.
  • [EXTERNAL_DOWNLOADS]: The skill uses ctx_fetch_and_index to download content from the web for indexing. The instructions mandate explicit approval before performing network fetches and warn against indexing sensitive data like credentials or tokens.
  • [INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because it indexes tool outputs and web content into an FTS5 database for later retrieval.
  • Ingestion points: Untrusted data enters the system through ctx_index (indexing session/tool output) and ctx_fetch_and_index (fetching external web pages).
  • Boundary markers: The instructions do not specify technical delimiters for the indexed content but provide strong behavioral warnings to the agent.
  • Capability inventory: The skill has high-impact capabilities, including arbitrary command execution (ctx_execute), file reading (ctx_execute_file), and configuration writes (ctx_upgrade).
  • Sanitization: There is no evidence of sanitization for indexed content; however, the skill's architecture emphasizes human-in-the-loop approval for all actions triggered by the retrieved context.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 13, 2026, 11:08 PM