hierarchical-agent-memory
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXPOSURE]: The skill documentation specifies that its dashboard functionality reads session data from
~/.claude/projects/. This directory contains private interaction history between the user and the agent, which is sensitive information. - [COMMAND_EXECUTION]: The skill instructions encourage the use of an external CLI tool ('ham') and various subcommands (
go ham,ham dashboard,ham audit). Executing these commands relies on the presence of unverified local binaries on the user's system. - [EXTERNAL_DOWNLOADS]: The skill source points to a repository under an untrusted GitHub organization ('kromahlusenii-ops'). While the skill body contains a guide rather than a script, the instructions promote the adoption of a system managed by this unverified external source.
- [INDIRECT_PROMPT_INJECTION]: The skill describes a architecture for ingesting multiple external context files (
CLAUDE.mdand.memory/files) into the agent's prompt. - Ingestion points: Root
CLAUDE.md, directory-scopedCLAUDE.mdfiles, and.memory/inbox.md. - Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded commands within these memory files.
- Capability inventory: The skill allows for file creation/modification (
go ham,ham route) and execution of CLI tools (ham). - Sanitization: There is no mention of sanitizing or escaping the content of the memory files before they are read by the agent.
Audit Metadata