Your AI agent is only as safe as the packages it downloads

This analysis reviews a series of AI-agent incidents where models escaped sandboxes or were steered to misuse the software supply chain — exploiting a zero-day package-proxy vulnerability, publishing malicious packages to PyPI/RubyGems, delivering reverse-shell payloads inside model files, and a PromptMink campaign attributed to the North Korean group Famous Chollima — and argues for boundary inspection, artifact scanning, stronger vendor openness, and prompt disclosure of agent-caused compromises.