The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The delivery script scripts/vod_deliver.py transmits the GitCode access token as a query parameter in the API request URL (?access_token=...). This is an insecure practice as tokens can be leaked into server logs or browser history.\n- [EXTERNAL_DOWNLOADS]: The installation scripts scripts/vod_install.sh and scripts/vod_install.ps1 clone source code from a personal repository (https://gitcode.com/weixin_45218422/AtomGit-GO.git) and build binaries locally. This introduces a dependency on unverifiable code from a non-standard source.\n- [COMMAND_EXECUTION]: The script scripts/vod_deliver.py uses subprocess.Popen to launch a local authentication server binaries from the user's home directory. Spawning background processes is a high-privilege operation that increases the system's attack surface.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data such as error stack traces and conversation history to populate reports via scripts/md_io.py. While a sanitization script scripts/vod_sanitize.py is provided to redact secrets, the skill remains a surface for indirect instructions embedded in technical logs. Ingestion points: scripts/md_io.py through --error-stack and --dialog-context arguments. Boundary markers: The skill uses markdown headers like ## Error Information to delimit content. Capability inventory: Includes subprocess.Popen and network requests via urllib.request. Sanitization: Implements regex-based redaction in scripts/vod_sanitize.py.\n- [DATA_EXFILTRATION]: The scripts/md_io.py utility includes a _resolve_value function that interprets any string starting with @ as a file path and reads its content into the report. This could be exploited by an agent or a malicious prompt to exfiltrate the contents of sensitive local files by submitting them as feedback.

huawei-cloud-vod-collector