memmachine
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates on long-term memory data which could potentially contain malicious instructions intended to influence the agent (indirect prompt injection). This is a standard risk for memory-management skills. The skill mitigates this surface by instructing the agent to resolve specific project/session scopes and requiring explicit user approval before any memory mutations (writes/deletes) or external service connections.
- Ingestion points: Memory payloads retrieved from the MemMachine server and configuration files.
- Boundary markers: The skill relies on structured instructions requiring the agent to pause for approval before acting on retrieved data.
- Capability inventory: The agent has access to
Bash,Read,Grep, andGlobtools. - Sanitization: The provided
preflight.pyutility verifies the integrity of the tool source code, though it does not perform runtime sanitization of individual memory payloads. - [SAFE]: The skill demonstrates a high security posture by including a defensive
preflight.pyscript designed to verify the local environment. This script implements several security controls: - Path Sanitization: It contains a blocklist of sensitive file names and stems (e.g.,
.ssh,.aws,.env,id_rsa,token,secret) to prevent accidental data exposure. - Secure Path Traversal: It uses
os.O_NOFOLLOWand directory file descriptors (dir_fd) to prevent symbolic link attacks during file verification. - Integrity Checking: It validates the source code against a manifest of SHA-256 hashes (
references/source.json) to ensure the agent is using verified versions of the integration logic. - Least Privilege: The instructions explicitly gate all potentially dangerous actions (installations, service startup, MCP registration, data deletion) behind mandatory human approval.
Audit Metadata