Two AI Labs Just Admitted What's Still Unsolved. Plan Like They're Right.

When the company that built the tool tells you, in its own words, that a core problem isn't solved, the useful response isn't relief that they're being honest. It's updating your plan to match what they just told you.

Two admissions, four weeks apart

In late October, Sam Altman and Demis Hassabis both said AI security is an unsolved problem days apart — not a marketing stumble, but the CEOs of OpenAI and DeepMind on record saying patching known holes isn't the same as the problem being fixed. Days later, Anthropic's draft IPO filing showed $16.3 billion in first-half revenue against an $80.7 billion net loss, with over $518 billion in compute commitments already signed. Neither event was new information about what AI can do. Both were the clearest admission yet, from the people with the least incentive to say it, that two things teams routinely treat as settled — "the model is secure enough" and "the price will hold" — are not settled at all.

This isn't the first version of either warning. Altman and Amodei briefed the UN Security Council on AI risk in person months earlier, and AI subscription pricing has been known to run below the actual cost of compute since at least May. What's new is the specificity — a named security failure mode instead of a general warning, an audited loss number instead of an analyst's estimate. Vague risk is easy to file away. A number on an S-1 is not.

What to do differently Monday morning

Not panic, and not switch vendors. Treat both admissions as inputs to two plans you should already have, and probably don't.

On security: stop assuming "the vendor handles it" covers you. If OpenAI and DeepMind's own CEOs say patching isn't the same as solving, the gap sits between what the model does by default and what your team actually sends it. That gap is yours to close, not theirs.

On cost: stop assuming this month's price is next year's price. A business losing $80.7 billion in six months while committing half a trillion dollars to future compute will eventually raise prices, change free-tier terms, or both. Knowing your number for "too expensive" before that happens is the difference between a planned switch and a scramble.

A compliance lead at a 60-person healthcare billing firm

She'd read the security story as a headline about OpenAI, not about her team — until she actually listed what her staff pasted into AI tools during a normal claims-review week. Patient IDs showed up twice, both times because someone was pasting a denial letter to get a faster rewrite, not because anyone decided that was fine. Nothing had leaked. But "nothing has leaked yet" and "we checked" turned out to be different claims, and she'd been treating them as the same one. She wrote a one-line rule — no patient identifiers in any AI tool, paraphrase first — and posted it where the team actually works, not in a policy doc nobody opens.

A solo bookkeeper running four client accounts through one AI subscription

He'd never priced a second tool, because the first one worked and switching sounded like a project he didn't have time for. After the IPO numbers made the rounds in a bookkeeping forum he follows, he spent twenty minutes pricing the same monthly workload on a cheaper model tier instead of a different vendor entirely — reconciliation summaries didn't need the top tier he'd defaulted to out of habit, and the cheaper one produced the same output. He kept his vendor. He just stopped paying for more model than four small accounts required, which bought him room if the price moves before he's ready to shop around.

The one thing

Altman, Hassabis, and Anthropic's own auditors aren't telling you to panic about security or cost. They're telling you neither is finished — which is useful information only if you act on it before the next admission arrives with your name on it.