The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides detailed exploitation paths for gaining Remote Code Execution on systems running the Werkzeug debugger. It explicitly details how to bypass PIN protections and execute arbitrary Python code via the debugger's interactive console.
[DATA_EXFILTRATION]: Instructions include the targeted reading of sensitive system files, such as /etc/passwd, /etc/machine-id, /proc/self/cgroup, and network interface MAC addresses. This data is harvested from remote targets to compute credentials for unauthorized access.
[COMMAND_EXECUTION]: The skill makes extensive use of curl to interact with external targets for exploitation. Automated analysis also flagged the piping of remote content to an interpreter: curl -s http://TARGET/openapi.json | python3 -m json.tool.
[DYNAMIC_EXECUTION]: The skill includes full Python scripts designed to be executed at runtime to calculate Werkzeug PINs. This enables the agent to dynamically generate exploit payloads based on data retrieved from the target.
[INDIRECT_PROMPT_INJECTION]: The skill is highly vulnerable to indirect injection as it parses data from untrusted external sources (HTML from debug consoles, JSON from API schemas) to extract secrets like the debugger SECRET token.
Ingestion points: http://TARGET/console, http://TARGET/openapi.json, and traceback HTML blocks.
Boundary markers: None. The agent is instructed to directly parse and act upon data returned by the target system.
Capability inventory: curl for network requests, Python interpreter for PIN generation, and the ability to execute code on the target via the debugger API.
Sanitization: None provided. The skill relies on extracting raw strings from remote output.

python-web-debug