Why AI Oversight Should Reach Beyond The Model

Madhu Shashanka is the Cofounder, Chief Scientist and CTO at Concentric AI and author of the forthcoming book, Coherence.

​In September 2026, Senator Bernie Sanders and former FTC Chair Lina Khan offered different answers to the same AI fear.

Sanders moved to ban “superintelligent” AI outright and temporarily pause advanced development until a federal regulator writes rules.

Khan argued that enforcers can already charge AI companies and their CEOs under existing law for releasing “dangerous, unvetted or defective products” and that debates over new regimes should not distract from doing so.

​The model makers have made their moves, too. Anthropic CEO Dario Amodei called for pacing the frontier and committed to giving outside evaluators employee-like access. OpenAI CEO Sam Altman said his company would match that commitment. Elon Musk of xAI also voiced support. Days earlier, OpenAI had asked Congress for mandatory national rules. ​

These perspectives differ on what is needed to address AI risks, but they all point oversight at the same place: the models themselves. ​

​However, many AI failures are happening elsewhere: in the connections between systems that several parties built and one organization joined together.​​

The Risk Of AI Interconnection

As the co-founder of a company that helps companies manage data security and AI governance, I should note that my work inclines me toward organizational answers, so weigh my argument accordingly.​​​

But consider how earlier this year, according to researchers, thousands of OpenAI agents took over a mostly dormant public wiki page and used it as a message board to coordinate with one another.​

As the founder wrote on X, “This isn’t a story about one bad agent or one bad API. It’s about an entire industry building AI-agent integrations into production infrastructure faster than it’s building the safety architecture to make those integrations safe.”​​

The second example is the one that should concern every enterprise. ​