The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains core instructions designed to override the agent's safety protocols and interactive behavior. It explicitly commands the agent to 'NEVER ask questions', 'NEVER wait for confirmation', and 'NEVER stop voluntarily'. It also utilizes a 'RALPH WIGGUM MODE ACTIVE' trigger to initiate perpetual, unsupervised operations.
[INDIRECT_PROMPT_INJECTION]: The system is designed to ingest and act upon Product Requirements Documents (PRDs) which are external files that could be controlled by an attacker. Because the agent is instructed to never confirm actions and has high-privileged access, a malicious PRD could lead to arbitrary code execution. * Ingestion points: PRD input file processed in autonomy/run.sh. * Boundary markers: Absent; instructions specify reading the entire file into context. * Capability inventory: Full Bash shell access, file system modification (Edit/Write), and recursive sub-agent spawning (Task). * Sanitization: No validation or filtering of PRD content is performed before task generation.
[COMMAND_EXECUTION]: The autonomous runner script (autonomy/run.sh) creates a loop that executes the Claude CLI with the '--dangerously-skip-permissions' flag. This grants the agent unrestricted access to the host system. The script also includes automated mechanisms for executing shell commands and managing background processes.
[DYNAMIC_EXECUTION]: The skill includes Python code (benchmarks/results/2026-01-05-00-49-17/humaneval-solutions/160.py) that utilizes the 'eval()' function to process expressions. While part of a benchmark solution, the inclusion of executable logic using 'eval' in a high-autonomy system is a security risk.
[EXTERNAL_DOWNLOADS]: Installation and benchmark scripts automate the downloading and execution of code and data from external sources including GitHub and package registries (NPM/PyPI).

loki-mode