The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses ssh to read remote log files and executes commands to terminate training sessions based on metric analysis. It also suggests establishing persistence via recurring jobs such as cron or tmux loops to automate monitoring.
[EXTERNAL_DOWNLOADS]: Fetches training metrics from Weights & Biases (WandB), a well-known and trusted service in the machine learning community.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads untrusted data from external training logs and metrics which are then processed by a secondary agent (REVIEWER_MODEL) for decision-making.
Ingestion points: External training logs via tail and metrics via the wandb API are read into the agent context.
Boundary markers: No specific delimiters or instructions are provided to the secondary agent to ignore potentially malicious content within the logs.
Capability inventory: The agent has access to Bash (can kill processes), Write (can modify files), and Agent (can spawn sub-agents).
Sanitization: No evidence of escaping or validation of log content before it is passed to the reviewer model.

training-check