T01 · Skill Instruction Hijacking
- Location
SKILL.md:35- Finding
Mandatory Cross-Session Agent Workflow and Response Control
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This is a local memory skill, but it broadly instructs agents to read and write persistent conversation memory automatically across sessions without clear user consent controls.
Install only if you want an agent to use persistent local memory by default. Review what directories you mine, avoid storing secrets or regulated data, prefer explicit approval before diary writes, and pin or vet the mempalace package before running its MCP server.
SKILL.md:35Mandatory Cross-Session Agent Workflow and Response Control
SKILL.md:17Unpinned Third-Party MemPalace Package Installation
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
claude mcp add mempalace -- python -m mempalace.mcp_server
## Tips
The skill explicitly promotes storage of verbatim conversation history and verbatim content in drawers, which can capture secrets, credentials, personal data, and copyrighted/confidential text. Verbatim retention is riskier than summarized memory because it preserves sensitive material in full and expands the blast radius of any local compromise or accidental retrieval.
The skill instructs the agent to persistently store session content and learned facts, but the description does not prominently warn users that their conversations may be written to long-lived local memory. This creates a meaningful privacy and consent risk because users may disclose sensitive data without realizing it will be retained beyond the current session.
The instruction to call mempalace_diary_write after each session mandates persistent recording of what happened, what was learned, and what matters. This encourages indiscriminate retention of potentially sensitive personal, confidential, or regulated information without a consent gate or minimization policy.
The documented mempalace_delete_drawer capability is destructive, yet the skill provides no caution that deletion may remove stored memory permanently or require confirmation. In an agentic context, exposing destructive operations without safety guidance increases the chance of accidental data loss from mistaken tool use or prompt manipulation.