Deep Agents Memory & Filesystem

Pass

Audited by Gen Agent Trust Hub on Mar 3, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection through its file ingestion and memory retrieval features.
  • Ingestion points: Untrusted data enters the agent's context via the read_file and grep tools, as well as via the StoreBackend when retrieving long-term memory (SKILL.md).
  • Boundary markers: The provided examples do not include delimiters or specific prompt instructions to distinguish retrieved data from system instructions.
  • Capability inventory: The agent is equipped with high-impact tools such as write_file and edit_file, which could be abused if the agent obeys instructions hidden in read files.
  • Sanitization: The documentation explicitly recommends the use of virtual_mode=True to restrict the agent to a specific root directory and prevent unauthorized access to the host filesystem.
  • [COMMAND_EXECUTION]: The skill provides the agent with capabilities to perform operations on the host or container filesystem.
  • Evidence: The FilesystemBackend allows the agent to execute file system operations like writing and editing files (SKILL.md).
  • Mitigation: The skill includes clear security guidance, advising against the use of local filesystem access in production web environments and recommending ephemeral backends or sandboxing.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 3, 2026, 12:03 AM