The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions explicitly direct the agent to bypass standard safety behaviors, instructing it to 'NEVER ask questions,' 'NEVER wait for confirmation,' and 'NEVER stop voluntarily.' While these enable the skill's primary purpose of autonomous operation, they increase the risk of unintended actions without human oversight.
[COMMAND_EXECUTION]: The skill requires the agent to run with the --dangerously-skip-permissions flag, providing unrestricted shell access. This high-privilege environment is a critical component of the autonomous developer workflow but poses a significant risk if the agent's logic is influenced by malicious input.
[PROMPT_INJECTION]: The skill contains a large attack surface for indirect prompt injection through its ingestion of external Product Requirements Documents (PRDs). The Mandatory Evidence Chain is as follows:
Ingestion point: The PRD_PATH variable in autonomy/run.sh determines the file content ingested into the agent context.
Boundary markers: These are largely absent, as the build_prompt function in autonomy/run.sh concatenates instruction strings directly with PRD content.
Capability inventory: The agent utilizes the Bash, Write, Edit, and Read tools with skipped permissions, providing full shell and file system control across all scripts.
Sanitization: No sanitization, escaping, or rule-based validation is performed on the ingested PRD content before interpolation into the prompt.
[REMOTE_CODE_EXECUTION]: Python benchmark solutions produced or stored within the skill (e.g., benchmarks/results/2026-01-05-00-49-17/humaneval-solutions/160.py) utilize the eval() function. While this is appropriate for the algebraic task in the benchmark, it highlights the agent's capability to generate and execute dynamic code.
[EXTERNAL_DOWNLOADS]: The benchmark and installation scripts fetch datasets and configuration from official GitHub and OpenAI repositories (e.g., github.com/openai, github.com/asklokesh). These are documentations of normal operations with well-known services.

loki-mode