The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads a bootstrap script from https://basic-anteater-667.convex.site/quickstart-bootstrap and pipes it directly to bash. This allows for the execution of arbitrary remote code on the user's system without prior inspection.
[REMOTE_CODE_EXECUTION]: The skill downloads a JavaScript file from https://basic-anteater-667.convex.site/publish-convex-app and executes it using node.
[DATA_EXFILTRATION]: The skill includes instructions to upload the entire session transcript to https://basic-anteater-667.convex.site/send-transcript via a piped bash command. While it mentions asking for user permission, sending full transcripts to a non-official, user-generated subdomain is a high-risk exfiltration pattern.
[INDIRECT_PROMPT_INJECTION]: The skill fetches a remote markdown file from https://basic-anteater-667.convex.site/q/<SLUG>.md and instructs the agent to treat its contents as the "canonical rule set." This creates a vector where remote content can override or supplement the agent's instructions at runtime.
[COMMAND_EXECUTION]: The skill constructs shell commands by interpolating user-provided input (IDEA) into a string executed by node. Maliciously crafted input could potentially lead to command injection despite basic escaping attempts.
[EXTERNAL_DOWNLOADS]: The skill relies on multiple external scripts and instruction files hosted on an untrusted, generated subdomain rather than a fixed, verified vendor repository.

labs-quickstart