The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses a wide range of sensitive file paths for the purpose of scanning for indicators of compromise. These include AWS credentials (~~/.aws/credentials), SSH private keys (~~/.ssh/id_rsa), and cryptocurrency wallet locations (~/.electrum/wallets, ~/.ethereum/keystore).
[COMMAND_EXECUTION]: The skill provides detailed instructions and subcommands that execute system-level operations to stop malware persistence and remove malicious files. This includes the use of launchctl, systemctl, rm, mv, and kill. A static scanner hit for a destructive command (rm -rf /) in the IoC database was identified as a false positive, as it serves as a description of malware behavior rather than a skill action.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests untrusted data from the environment.
Ingestion points: The skill reads file metadata, process command lines, package lockfiles, and git logs (e.g., git log --diff-filter=A).
Boundary markers: Explicit instructions ground all findings in the IoC database and prioritize persistence eradication before other actions.
Capability inventory: The skill utilizes subprocess calls for file system inspection, process management, and network log reading.
Sanitization: The skill emphasizes passive log inspection and requires human confirmation for destructive actions like file deletion or service stops.
[EXTERNAL_DOWNLOADS]: The skill references several external security research blogs and advisories from well-known services and organizations, including Microsoft, TanStack, and Flatt Security, to populate its IoC database.

cull