The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/nettrace.py uses subprocess.run to invoke the adb CLI. It implements safe command construction by passing arguments as a list and using shlex.quote for user-controllable inputs, preventing local command injection.\n- [DATA_EXPOSURE]: The tool captures network traffic which can contain PII. The mitm_capture.py script includes a redaction list for sensitive headers like Authorization and Cookie to mitigate the risk of credential exposure in saved logs.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, specifically captured PCAP files and APK binaries (scripts/nettrace.py, scripts/mitm_capture.py). While it performs parsing and extraction, the resulting text (e.g., hostnames, URLs) is fed back to the AI agent. This creates a surface for indirect prompt injection if the application being analyzed contains malicious strings designed to influence LLM behavior. Evidence chain: Ingestion at parse_pcap/extract_endpoints; no boundary markers on output lists; capabilities include subprocess execution (adb); sanitization present via shlex.quote for command arguments.\n- [SAFE]: Temporary configuration files used to store proxy state are created using os.open with O_EXCL and restricted permissions (0o600), ensuring that other users on the system cannot access or manipulate the saved settings.

android-network-trace