The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface.
Ingestion points: The skill reads content from a target SKILL.md and user-provided test inputs (prompts and scenarios) to drive its optimization loop.
Boundary markers: Absent; instructions for the agent do not include delimiters to isolate the agent's instructions from potentially malicious content in the external skills or inputs.
Capability inventory: The skill possesses file system write access (overwriting SKILL.md), shell command execution (open dashboard.html), and the ability to trigger other skill executions.
Sanitization: Absent; content from external sources is processed and incorporated into prompts for the mutation logic without validation.
[PROMPT_INJECTION]: Autonomy and Oversight Suppression. The skill instructs the agent to 'NEVER STOP' and 'not pause to ask the user' once the loop begins, effectively removing human oversight and allowing autonomous modifications and executions across many experimental cycles.
[COMMAND_EXECUTION]: The skill executes the 'open' command to launch the user's browser for viewing the generated HTML dashboard.
[EXTERNAL_DOWNLOADS]: The generated dashboard is configured to fetch the Chart.js library from a public content delivery network (CDN).

autoresearch