Type: url
Status: generated
Summary
GitHub’s account of the security guidelines behind its hosted AI agents, written from the premise that the more agentic a product is the greater the chance and impact of it going off its guardrails, losing alignment or being manipulated. It names three risk classes — data exfiltration, impersonation and action attribution, and prompt injection — and six rules applied across GitHub’s hosted agentic products: make all context visible and strip hidden Unicode or HTML directives; firewall the agent’s network access; withhold sensitive information such as CI secrets and revoke tokens after a session; prevent irreversible state changes without a human in the loop; attribute actions to both initiator and agent; and gather context only from users with write access.