The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard shell utilities (awk, grep, sed, comm, sort) to process browser snapshots and calculate diagnostic grades. These operations are performed on temporary files within /tmp and do not involve unauthorized system modifications or persistence attempts.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it processes and interprets content from external, untrusted web pages.
Ingestion points: Content is ingested from the browser via agent-browser snapshot and agent-browser get url into shell variables and temporary files (SKILL.md).
Boundary markers: No explicit delimiters or instructions are used to distinguish untrusted webpage content from the agent's operational logic.
Capability inventory: The skill utilizes capabilities such as agent-browser click, agent-browser eval, and agent-browser cookies set, which could be manipulated if an attacker can influence the snapshot or URL data.
Sanitization: While the skill uses quoted shell variables and sanitizes certain identifiers (like ref tags) via sed, there is no comprehensive validation of the content retrieved from the web page before it is processed.

diagnose-instant-nav