wordcraft

Fail

Audited by Gen Agent Trust Hub on Oct 22, 2026

Risk Level: HIGHCREDENTIALS_UNSAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill provides explicit instructions for the agent to access and read a local authentication key file required for communication with the application's control channel.
  • Evidence: The skill specifies the path ~/Library/Application Support/WordCraft/control-key.local-<port> and provides the command cat ~/Library/Application\ Support/WordCraft/control-key.local-7981 to retrieve the secret.
  • [DATA_EXFILTRATION]: The skill exposes and instructs the agent to interact with sensitive application-specific file paths containing user profile data and document recovery snapshots.
  • Evidence: Mentions paths to the ~/Library/Application Support/WordCraft directory and the recovery/ folder.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external and untrusted document formats (including .docx, .html, .rtf, and .md) and possesses significant system capabilities without explicit safety controls.
  • Ingestion points: Document ingestion occurs via open_document, wordcraft-cli convert, and various format-specific readers in crates/engine/src/io.rs.
  • Boundary markers: Absent; there are no instructions to use delimiters or warnings for the agent to ignore potentially malicious embedded instructions in documents.
  • Capability inventory: The MCP execute tool and CLI run --cmd flag allow the execution of any internal engine command, including file system operations (file.save) and user interface manipulation.
  • Sanitization: No sanitization of document content is mentioned before processing.
  • [REMOTE_CODE_EXECUTION]: The skill documents procedures for building and running the application and its CLI directly from an external source repository.
  • Evidence: Usage recipes include cargo run --release -p wordcraft-cli and cargo run ... --example ui_shot targeting the remote source code.
  • [COMMAND_EXECUTION]: The skill exposes a comprehensive internal command registry to the agent, enabling a wide range of application and system-level actions.
  • Evidence: Describes the command registry at crates/engine/src/cmd/mod.rs::registry and allows invocation via automation methods like engine.execute.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 22, 2026, 01:11 PM