The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions to execute a sudo command that modifies the ptrace_scope kernel parameter. This configuration change reduces system security by allowing processes to trace and inject code into other processes owned by the same user. Evidence: echo 0 | sudo tee /proc/sys/kernel/yama/ptrace_scope in SKILL.md.
[COMMAND_EXECUTION]: The skill relies on various shell commands for environment setup and network interaction. Evidence: Use of nc (netcat) to establish connections to local debug ports (nc 127.0.0.1 4444) and custom environment build tools (pm.build_env).
[DYNAMIC_EXECUTION]: The skill facilitates arbitrary Python code execution through integrated debugging tools like pdb. Evidence: The !stmt and interact commands allow the agent to execute arbitrary Python code and manipulate the state of the running application.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by exposing the agent to untrusted data from debugged processes. 1. Ingestion points: Agent reads debugger output including variable values and stack traces in SKILL.md. 2. Boundary markers: No delimiters or warnings are used to distinguish debugger output from instructions. 3. Capability inventory: The skill enables arbitrary command execution and Python code execution via debugger interfaces. 4. Sanitization: No sanitization or filtering is applied to the output of the debugger before processing by the agent.

python-debugpy