The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with external, untrusted web content while holding an authenticated session, creating a surface for indirect prompt injection where malicious instructions on a webpage could attempt to influence the agent.
Ingestion points: The agent reads and processes untrusted HTML and page data from arbitrary websites (referenced in SKILL.md).
Boundary markers: The instructions mandate setting "destructive boundaries in every goal" to restrict the agent's actions.
Capability inventory: The agent possesses the ability to perform authenticated actions such as clicking, submitting forms, and deleting data within the browser context.
Sanitization: The skill requires the agent to confirm high-risk actions with the user and to identify and report attempts by web content to override instructions.
[SAFE]: All external URLs (agent.tinyfish.ai, docs.tinyfish.ai) are official resources belonging to the skill's vendor infrastructure.
[SAFE]: The skill incorporates robust defensive measures to prevent credential theft, specifically prohibiting the agent from reading sensitive local files like .env or SSH keys, and forbidding the use of passwords in goal descriptions.

tinyfish-authenticated