Rivals talk AI safety as risk warnings mount and shadow AI spreads unchecked

There's a strange irony sitting at the center of today's AI news: the same three companies racing each other into oblivion or riches — OpenAI, Google, and Anthropic — are apparently talking to each other about how to keep things safe. Sources say the talks have been going since July, kicked off by Demis Hassabis's proposal for a U.S.-led AI "Standards Body." A spokesperson confirmed the conversations exist but wouldn't say much else, which tells you almost everything you need to know about where this stands. Competitors coordinating on safety while still competing for the same enterprise contracts and the same scarce compute is a genuinely hard problem, and I'm skeptical any voluntary standards body survives contact with a funding round.

That skepticism matters more given the political backdrop. Talk of AI "existential risk" has moved from fringe Discord servers into mainstream political discourse, and the pressure for real regulation is building in response. Trump, though, keeps rejecting anything that looks like a brake on development, and has taken specific aim at Anthropic — a company whose entire branding rests on being the safety-conscious lab. That's a genuinely odd position for an administration to occupy: dismissing existential risk concerns while the researchers raising them keep multiplying. Bilal Chughtai, who left Google DeepMind in July, is the latest to go public with warnings about where this technology could lead. I don't think every one of these researcher exits should be read as a five-alarm fire, but the pattern is now long enough that treating it as noise feels increasingly lazy.

Meanwhile, the practical safety problem isn't superintelligence — it's Bob in accounting running an unauthorized AI agent to sort his invoices. Imprivata's new report found that 72% of healthcare organizations have unapproved AI tools running somewhere inside their walls, and 88% are bracing for autonomous agents that will need proper identity-based governance. Healthcare is a useful bellwether here because it combines high regulatory stakes with genuinely chaotic IT environments, and if three-quarters of hospitals already have shadow AI, I'd wager the number is similar or worse in finance, law, and government. This is the unglamorous side of the safety conversation — not "will AI end civilization" but "does anyone actually know what's touching patient records right now." Microsoft clearly sees the opportunity: its new Microsoft 365 G7 bundle is explicitly built around agent governance and compliance for government agencies, betting that the path to safe AI adoption runs through access control, not philosophy.

By the way, Boston Dynamics quietly shipped Spot 5.2 this week, giving the robot better autonomous fault-detection in industrial settings. It's a small, almost boring update next to talk of existential risk — and that contrast is the point. The AI safety conversation happening in op-eds and Senate hearings is about models nobody has built yet, while the one happening in hospital IT departments is about tools already running unsupervised today. I suspect the second conversation will shape real policy faster than the first.