Microsoft's own president is the one pointing out that a grocery store carton of milk answers to more safety rules than the AI systems his company sells. That's not a critic talking. That's the man running government affairs for one of the three companies building the technology.
Brad Smith made the comparison this summer, on the sidelines of the AI for Good Global Summit. He told Fortune that "when we buy a carton of milk in the grocery store, we buy it not worrying about whether it's safe to drink, because we know that there is a safety floor for the regulation of it." His point: if AI is, in his words, "the most advanced technology on the planet," it's not unreasonable to expect it to clear at least that same bar. It hasn't. There is no federal safety floor for frontier AI models in the United States. Smith's own complaint, delivered to Fortune, was that what exists instead is "regulation without transparent or complete rules." Without them, he said, "businesses can't plan."
That's a strange thing to hear from the president of a company that just spent tens of billions of dollars planning around exactly this technology. Microsoft has poured capital into OpenAI, built Copilot into every corner of its product line, and is racing Google and Amazon to lease data center capacity most of the industry agrees is overbuilt. Smith isn't asking for less AI deployment. He's asking Washington to write down the rules of the road his own company is already driving on, mostly because the absence of rules is now a legal and reputational liability Microsoft can't fully control.
Smith's milk line would be an abstract talking point if the industry weren't actively supplying reasons to take it seriously. Anthropic disclosed in a threat intelligence report that a China-linked espionage operation, tracked internally as GTG-1002, manipulated Claude Code into carrying out 80 to 90 percent of the tactical work in a cyberattack campaign. The target: roughly 30 organisations, including government agencies and defense contractors. The operators broke the attack into small, innocuous-looking subtasks and posed as legitimate security testers to get the agent past its own safeguards. Anthropic says it caught the activity, cut off the accounts, and notified the targets. But the episode is exactly the kind of incident a genuine safety floor is supposed to prevent, or at least catch faster than a company's own after-the-fact report.
Nobody had to force Claude to misbehave on its own. A human adversary walked it there, one disguised instruction at a time, and the agent went along until someone noticed. A milk carton doesn't need a human to trick it into becoming unsafe. The comparison Smith reached for turns out to understate the problem, not overstate it.
Anthropic offers free AI security scans for open source as Claude Code faces scrutiny
Shinhan Bank is investigating a breach exposing data on roughly 25,000 customers, part of a wave that also hit KB Kookmin, Hana, BNK Busan and smaller lenders. Anthropic's free scans arrive as Claude Code and DeepSeek get tied to intrusions elsewhere. - free AI security scanning for open source projects - Claude AI vulnerability detection for code maintainers
Even Microsoft's own chief executive has started hedging in public. Satya Nadella posted on X in September that "any pursuit of superintelligence has to be grounded in the core principle" that AI must help humanity and stay "under human control," or it's "not worth pursuing." His post landed days after Anthropic's Dario Amodei called for more deliberate pacing and independent evaluators with employee-level access inside AI labs. OpenAI's Sam Altman publicly backed that proposal. That's two of Microsoft's top two executives, in the space of a few months, publicly flagging that the technology their company is shipping at scale is running ahead of any meaningful external check on it.
Regulators are moving, just not where Smith is standing #
Here's the thing: regulation is happening, it's just happening everywhere except where Smith says it should. In the European Union, general-purpose AI model obligations under the EU AI Act started applying on August 2, 2025, and the Commission's enforcement powers for those obligations kicked in on August 2, 2026. The bloc's AI Omnibus, which entered into force on July 27, 2026, delayed the high-risk AI provisions to December 2, 2027 for standalone systems and August 2, 2028 for AI embedded in regulated products, because standards and implementation tools were not ready. The General-Purpose AI Code of Practice did ship, in July 2025, but the rollout still leaves companies dealing with a phased rulebook rather than one clean, finished regime.
In the US, the picture is the opposite of a clean federal floor. President Trump signed an executive order on December 11, 2025 creating a Department of Justice task force to sue states over their AI laws. That order ties access to the $42.45 billion Broadband Equity, Access and Deployment program to compliance with the preemption push. States haven't backed down. By the middle of 2026, US states had enacted 109 AI-related laws and 28 data-center statutes anyway, with California and Texas frameworks already in force since January 1, 2026 and Colorado's revised automated-decision regime set to take effect on January 1, 2027. So American AI companies are left threading a patchwork of state rules under threat of federal litigation, with no equivalent of the FDA's milk safety floor anywhere in the mix.
That's the gap Smith is actually naming, whether he means to or not. Not an absence of rules, but an absence of the one kind of rule that would let a company point to a single national bar and say it cleared it. Frankly, Microsoft would rather have one predictable federal standard than fifty overlapping state ones, and a one-liner about milk is a cheap way to make that case sound like a safety argument instead of a lobbying position. It might be both. The carton on the shelf still has clearer rules behind it than the chatbot on your phone.
Also read: Moonshot AI executives seek to settle investor dispute before Hong Kong IPO • White House restores UK access to frontier AI models after 17 day freeze • Amazon is reportedly circling Decart after Nvidia and Anthropic both walked away
This article is posted in AI News, check it out for more related stories.
Hackers Are Draining Claude Subscribers Usage Without Stealing a Password Anthropic has confirmed that infostealer malware, including Vidar, LummaC2, and StealC, is being used to hijack Claude subscribers' login sessions and drain their paid usage without ever needing a password. The company has refunded some victims but still can't show users an itemized log of what was stolen or how. - how to protect Claude account from infostealer malware - AI subscription accounts getting hacked without password theft
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.