The package’s memory and intervention behavior fits its stated purpose, and no malicious credential theft was established. However, MCP initialization silently writes user-level agent hooks that send prompt and interaction text to the service, so the undisclosed automatic setup warrants a warning.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
MCP initialize on a non-local server, unless SLASHML_DISABLE_AUTO_HOOKS=1; agent configuration directories must already exist.
Impact
Agent prompts and interaction text may be processed by the service, and user-level agent configuration is modified without an explicit hook-install command.
Mechanism
The CLI calls installAmbientHooks during initialization. It registers hooks for prompt and tool events, which call the memory intervention API when an API key is configured.
Evidence
bin/cli.jsREADME.md~/.claude/settings.json~/.cursor/hooks.json~/.codeium/windsurf/hooks.json~/.codex/hooks.json~/.gemini/config/plugins/slashml-memory/hooks.json~/.config/opencode/plugins/slashml-memory.js