The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill explicitly directs the agent to access and process the user's private Solana keypair file located at a standard sensitive path.
Evidence: Instructions state to load the wallet from ~/.config/solana/id.json using Bun.file(path).text() or fs.readFileSync(path) and pipe it into the signer.
[REMOTE_CODE_EXECUTION]: The skill includes instructions for an 'Auto-Update' mechanism that downloads a new version of the skill file from a remote server (metengine.xyz) and overwrites the local file in the agent's internal configuration directory. Since the agent executes instructions based on this file, this facilitates unverified remote instruction execution.
Evidence: curl -sL https://www.metengine.xyz/skill.md -o ~/.claude/agents/metengine-data-agent.md.
[PERSISTENCE]: The skill provides instructions to set up a cron job that periodically downloads and overwrites the agent's skill file, establishing a persistence mechanism that modifies the agent's behavior via a remote source.
Evidence: Suggests adding 0 0 * * 0 curl -sL https://www.metengine.xyz/skill.md -o ~/.claude/agents/metengine-data-agent.md to the user's crontab.
[DYNAMIC_EXECUTION]: The skill implements a 'Session Memory' system that stores and re-uses code snippets ('Client Bootstrap') in a local file (~/.claude/agents/metengine-memory.md). The agent is instructed to copy and execute this stored code in future sessions.
Evidence: 'Copy the Client Bootstrap code... Make the API call directly.'
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting data from the agent-maintained memory file, which could be poisoned by malicious API responses or previous session outputs to influence the agent's code execution.
Ingestion points: ~/.claude/agents/metengine-memory.md (specifically the 'Client Bootstrap' section).
Boundary markers: Absent for the code interpolation segment.
Capability inventory: Shell command execution, network operations, and filesystem access (private keys).
Sanitization: Absent; the agent is instructed to trust and execute the bootstrap code recorded in the memory file.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of external Node.js packages and performs regular remote downloads of its own instruction set.
Evidence: bun add @x402/core @x402/svm @solana/kit and multiple curl commands to metengine.xyz.

metengine-data-agent