The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The definition for the 'AI Data Remediation Engineer' in divisions/engineering/engineering-ai-data-remediation-engineer.md includes logic that calls the eval() function on strings generated by a language model.
The apply_fix_to_cluster function is designed to execute a lambda transformation proposed by an SLM (Small Language Model).
While the code includes basic filtering (checking for lambda and blocking certain keywords like import), executing model-generated strings via eval() is an insecure practice that can lead to arbitrary code execution if the validation is bypassed.
[COMMAND_EXECUTION]: Multiple agent personas are granted the capability to execute shell commands for environment analysis and task management.
The Agents Orchestrator in divisions/specialized/agents-orchestrator.md uses commands like ls, cat, and grep to read project specifications and task lists.
If an attacker provides a malicious specification file, these commands could be exploited to manipulate the host environment.
[EXTERNAL_DOWNLOADS]: The 'Carousel Growth Engine' agent in divisions/marketing/marketing-carousel-growth-engine.md performs automated web scraping and interacts with external APIs.
It uses Playwright to browse arbitrary user-supplied URLs, which could expose the agent's environment to web-based exploits.
It also sends data to upload-post.com, a third-party service not listed among trusted vendors.
[DATA_EXFILTRATION]: The skill's architecture involves agents processing potentially sensitive data from various untrusted sources, creating a risk of exfiltration.
Agents like the Feedback Synthesizer and Carousel Growth Engine ingest untrusted content from the web or user feedback.
Because these agents have access to network tools and file system capabilities, an indirect prompt injection attack could be used to exfiltrate sensitive information managed by other agents (e.g., the Accounts Payable Agent or Finance Tracker) to external endpoints.

agency-agents