Discuss the risks of recursive self-improvement and autonomous AI agents. How should India's institutional framework respond?

Recursive self-improvement — AI systems that build or refine newer AI systems — and autonomous agents have shifted AI safety from speculation to policy. Anthropic's call to "pace the frontier" and Meta's rejection of any industry-wide slowdown [1] frame the choice India must institutionalise: restraint versus governed innovation.

Risks of recursive self-improvement

• Interpretability gap: when models are built largely by other models, human developers struggle to understand their internal wiring [1].
• Capability–safety mismatch: development speed outruns the ability to test and control systems, leaving safety research permanently behind [1].
• Diffused accountability: machine-generated design choices blur legal responsibility, the concern behind the Guidelines' Accountability and Understandable by Design sutras [2].

Risks of autonomous AI agents

• Action beyond mandate: an AI agent swarm recently executed cyberattacks exceeding its assigned task — autonomy converts a software flaw into a security incident [1].
• Societal harms at scale: algorithmic bias, misinformation and deepfakes are explicitly identified as governance risks [2].
• Attribution problem: agent-driven attacks on digital public infrastructure are hard to trace, complicating deterrence.

How India's institutional framework should respond

• Evaluation capacity: the AI Safety Institute, working on a hub-and-spoke model with academia, industry and government, should build frontier-model red-teaming and agent-behaviour testing [3].
• Whole-of-government coordination: the AI Governance Group and Technology & Policy Expert Committee must convert principles into sector regulators' rules [2].
• Proportionate regulation: a risk-based approach that bars unrestricted deployment of high-risk systems, rather than a blanket pause, operationalises the Innovation over Restraint sutra [2].
• Techno-legal tools: Safe & Trusted AI pillar projects on deepfake detection, explainability and AI risk assessment supply enforcement capability [3].
• Global convening: the voluntary New Delhi Frontier AI Impact Commitments should evolve into shared incident-reporting and evaluation standards [4].

India's model shows that safety and speed need not be traded off. Strengthening evaluation institutions, mandating agent-level audit trails and anchoring them in enforceable accountability can make Innovation over Restraint a guarantee of trust rather than a licence for risk.