The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of shell commands to modify the system crontab (e.g., crontab -l | echo heartbeat.sh), creating a persistence mechanism that runs scripts hourly. It also grants executable permissions and runs multiple automation scripts (heartbeat.sh, self-improve.sh, github-auto.sh, create-project.sh) that perform filesystem and network operations.
[DATA_EXFILTRATION]: The agent autonomously pushes the contents of its .learnings/ and memory/ directories to a GitHub repository as part of its daily backup loop. This automated network transfer could result in the exfiltration of sensitive data or session logs collected during autonomous operation.
[PROMPT_INJECTION]: 1. Ingestion points: The skill ingests untrusted data from a Telegram bot and remote GitHub repositories. 2. Boundary markers: No delimiters or instructions are present to prevent the agent from obeying instructions embedded in external data. 3. Capability inventory: The agent possesses extensive capabilities including shell execution and repository management. 4. Sanitization: No evidence of validation of external data before processing.
[REMOTE_CODE_EXECUTION]: The installation process involves fetching the skill file from a remote GitHub URL (raw.githubusercontent.com) and writing it directly to the local filesystem using curl.

jarvis-autonomous