The Agent Skills Directory

[DATA_EXFILTRATION]: The skill performs broad read operations across sensitive system locations to identify anomalies. It accesses user LaunchAgents (~/Library/LaunchAgents), enumerates all hidden files and directories in the user's home directory, and reads the user's crontab. While intended for malware detection, this capability allows for significant data exposure of local system configuration and user patterns.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources which could be manipulated by a malicious actor already present on the system.
Ingestion points: System process lists (ps), LaunchAgent plist contents, and file paths in ephemeral directories.
Boundary markers: None specified to prevent the agent from misinterpreting malicious strings inside these data sources.
Capability inventory: The skill uses shell commands and the codesign utility to inspect files.
Sanitization: No explicit sanitization of extracted strings (like paths from plists) is mentioned before they are used in subsequent shell commands.
[COMMAND_EXECUTION]: The skill relies on executing a wide array of shell commands (awk, grep, find, codesign, scutil) to perform its heuristics, which is standard for system administration tasks but increases the overall attack surface if user input or untrusted file contents were to influence the command arguments.

malware-scan