New safety guidelines released for advanced artificial intelligence systems in India

The Indian Computer Emergency Response Team has issued new security rules for artificial intelligence systems that work on their own. These rules apply to systems that can plan tasks, use outside tools, and pass work to other artificial intelligence programs.

Security Checks

Auditors must test if these programs can use tools without permission, exceed their assigned limits, or misuse login details. The framework also requires checks on emergency-stop buttons and how agents handle memory from past tasks. Organisations must test if private user data can leak or if memory can be tampered with.

Multi-Agent Rules

When multiple programs work together, each one needs a clear identity and set permissions. All messages between them must be logged and checked to prove who made a request. Companies must also stop endless loops where programs keep assigning tasks to each other.

Context

These guidelines aim to prevent security breaches and data leaks as artificial intelligence tools become more independent and capable of acting without human supervision. Similar technical standards are often updated by cybersecurity agencies to keep pace with rapid software changes.