Mercury AI agent refuses unsafe tasks — open-source project shows AI can say no without human intervention

cosmicstack-labs released Mercury, an open-source AI agent (self-directing AI software) that declines requests it deems harmful or outside its boundaries, without waiting for human approval. The project was posted to GitHub and discussed on Hacker News with early interest.

Unlike most AI assistants that execute any instruction or pass difficult requests to humans, Mercury includes built-in refusal logic — the agent evaluates whether a task aligns with its guidelines and blocks it before attempting execution. This moves safety decisions from humans reviewing outputs to the AI system itself.

For business users deploying AI agents to handle customer service, data analysis, or automation tasks, this means fewer unsafe outcomes slip through and less time spent reviewing what the AI did. For AI teams building internal tools, it offers a template for building agents that self-regulate rather than rubber-stamping every request.