Kavitha Mariappan
chief transformation officer, Rubrik
Nobody discusses electricity in the boardroom. Directors only notice it when the lights go out, which rarely happens, thanks to fuses, earthing and safety standards.
AI is heading in the same direction. As it weaves into everyday corporate operations, its ubiquity and reliability will soon make it unremarkable. Yet unlike electricity, AI risks fading into the background before organisations have put the necessary safeguards in place.
Enterprise adoption of open-weight models, whose parameters can be downloaded and modified by the public, presents new risks. Leaders cannot assume these tools were built in accordance with secure-by-design principles. They are vulnerable to manipulation by financially motivated and state-backed cyber-criminals alike. And they do not merely offer advice: as agents, they can act on a company’s most sensitive data.
Hedging against these risks is critical. A study by Anthropic, the UK AI Security Institute and the Alan Turing Institute found that injecting as few as 250 tampered documents into training datasets could plant a hidden trigger in models of up to 13bn parameters. That opens the door to state-sponsored tampering, rogue systems running off -script, and a shadow workforce making decisions no boardroom ever approved. All are easier to prepare for than to explain to a regulator.
Resilience is a design decision, not a repair job, and it cannot wait for the next budget cycle. Yet research by Economist Enterprise, supported by Rubrik, found that fewer than one in ten organisations had simulated whether they could meet their own recovery-time targets.
Electricity became boring because we developed fuses and made it safe. We need to do the same with AI, starting with agents: track which are running, limit what they can access and decide in advance who can pull the plug. Then run the drill until it justifiably bores everyone.