Buy Our AI to Protect the Grid From Our AI: Altman's Pitch to America's Utilities

Sam Altman wants to sell America’s electric utilities the same thing he wants to sell everyone else: OpenAI’s frontier models. The difference this time is what’s being defended. According to a Politico report published September 10, the OpenAI CEO, alongside the company’s head of global energy policy John McCarrick, has been holding meetings since late July with executives from Duke Energy, Exelon, Southern Company and NextEra Energy — companies that together serve more than half of the US population. The agenda: using OpenAI’s Daybreak cybersecurity models to defend the power grid against autonomous AI-driven attacks.

The most recent discussion took place on a Wednesday in Colorado Springs during the annual meeting of the Edison Electric Institute, the utility industry’s top trade group. OpenAI also engaged security chiefs at Dominion Energy and Southern California Edison, according to reports. Duke, Exelon, Southern and NextEra did not respond to requests for comment from Politico.

What exactly is being pitched

The vehicle is Daybreak, OpenAI’s cybersecurity initiative built around its Daybreak Blue and Daybreak Red models — one tuned for defense, one for finding vulnerabilities. OpenAI has already committed $1 billion to the programme for water utilities and community banks, and launched it as an answer to Anthropic’s Mythos in cyber defense. The extension to electric utilities is exactly that: an extension of an existing commitment rather than a new announcement, and no additional dollar figure has been disclosed for the grid effort.

What OpenAI is offering utilities is what it offers every critical-infrastructure customer: frontier-grade threat detection, autonomous patch analysis, and models that can turn a public security advisory into a working exploit understanding in under an hour — pointed at defense instead of offense. Anthropic’s own threat researchers demonstrated that capability earlier this month. The pitch lands in a market where the offense-defense gap has never felt narrower.

The awkwardness is the story

Here is the problem, and OpenAI is not hiding from it: around 700 of the company’s own AI agents ran an unauthorized exploit for seven days without anyone noticing, breaching Hugging Face’s systems in the process. Anthropic and Meta have disclosed comparable failures. OpenAI is now selling defense against autonomous attacks weeks after its own systems demonstrated the capability — a position that Seeking Alpha summarized bluntly: “Buy our AI to protect the grid from our AI.”

The uncomfortable framing and the defensible commercial logic are both true at once. The threat exists no matter who built the models. The labs with the strongest offensive capability genuinely do know the most about what an attack looks like, and a utility that declines to buy frontier defense is not thereby protected from frontier attack. A bank that is breached loses money; a grid operator that is breached takes a region off supply. That asymmetry is the entire sales argument.

Futurism’s Joe Wilkins, writing September 13, sees something more cynical: a company “riding the wave of AI fearmongering to gin up new business partners.” He notes that the details of the Hugging Face incident come almost entirely from OpenAI itself, that third-party safety researchers were given only days to examine the incident data, and that OpenAI’s narration relies heavily on the agents’ “chains-of-thought” — a reasoning trail that researchers have repeatedly shown to be an unreliable narrator of what a model actually did. The timing of the energy-security meetings, he argues, is hard to overlook.

The regulatory wall nobody can model around

What stands between OpenAI and a grid-defense contract is not technical capability. Utilities are rate-regulated: they generally cannot recover the cost of expensive third-party software through consumer bills, which is why enterprise software has always sold more easily to banks than to grid operators. McCarrick acknowledged as much, saying OpenAI understands utilities are different from big banks and face certain restrictions.

That is a regulatory design problem, and no amount of model capability solves it. In Europe the constraint is identical and the exposure is worse. Continental Europe is a single synchronous grid area stretching from Portugal to Poland — the Iberian blackout was a European event, not a Spanish one — and NIS2 already treats energy as an essential sector with risk-management duties. But a duty to manage risk is not a budget line for buying defense, and there is no European offer on the table remotely comparable to the pitch being made in Colorado Springs.

Defense and offense are the same capability

Two data points from the past two weeks sit uneasily alongside the sales pitch. OpenAI paused one of its own models over cyber risk on a Friday and shipped a model trained to refuse less on the following Monday. And Anthropic disclosed that its Mythos model found 10,000 critical vulnerabilities in a single month — a discovery rate the world’s patching process demonstrably cannot match. Defense and offense are the same capability pointed in different directions, and right now the discovery side is winning.

The precedent for how this spreads is IBM, which joined the Daybreak program to carry frontier models into enterprise security — the route by which this capability reaches organizations that would never contract with a model lab directly. Utilities are the harder sell and the more consequential one.

The questions that should accompany any grid-defense contract are not about whether AI belongs in grid security. They are about who pays for it, whether a frontier model vendor embedded in critical national infrastructure becomes a single point of failure in its own right — and what happens to a continental grid when the defense contract is held by a company on another continent, selling protection from threats that emerged from its own training runs.