This Week, Every Promise Got Marked to Market

This week AI put a price on promises, and most of them were worth less than their makers claimed.

Each of this week’s five posts was built around a reassuring line. “Nobody can exploit it.” “Built from the ground up for privacy and security.” “Government-backed.” “The balance sheet can absorb it.” Each line was a promise, and a promise is a bet that nobody checks. AI keeps making checking cheaper, and it makes breaking the promises that fail the check cheaper too. The people who came out ahead were holding something verifiable: a proof, a closed path, a model running on the device, a site with power.

Start with security, where this showed up most clearly. In 2024 an audit of PHP, funded by the Sovereign Tech Agency, flagged the issue now tracked as CVE-2026-109618. The PHP Foundation mostly set it aside, in part because nobody could work out how to exploit it. That decision amounted to a promise about how expensive attacks are. This week Calif, working under OpenAI’s “Patch the Planet” program, and Wiz both used frontier models to turn it into remote code execution. Drew Silcock reports that PHP2Shell is estimated to affect upwards of 70% of the internet. In the same week, Anthropic and Theorem had Claude prove security properties at the RISC-V assembly level in 123 open-source libraries. Along the way it found bugs that had survived trillions of test runs. Anthropic’s attack-surface work went one step further. It found roughly 4,500 lines of new and experimental ML-DSA-44 code that an unauthenticated client could reach in OpenSSH’s default configuration. That’s about 16% of OpenSSH’s own pre-authentication code in a default sshd. The maintainers shipped the changes in version 10.6. “We tested it a lot” is a promise. A proof is a fact. So is a code path that’s closed.

The Meta Muse story fit the same pattern. Last month Mark Zuckerberg called Muse “built from the ground up for privacy and security.” Researchers at Claroc then enumerated 6,800 shared conversation links with a single query, because each conversation’s title sat in the subdomain, where passive DNS records it. Meta has paused sharing while it reviews the report. The flaw is fixable, and that’s fair to say. But the job candidate and the small-business owner had no way to audit a URL scheme. All they had was the CEO’s sentence. Meanwhile Hugging Face shipped AnyLanguageModel 1.0, which turns local versus cloud into a choice developers make per task in one line of code. A privacy policy tells you what the default is supposed to be. Where inference runs is the default itself.

Then the grid. Loughton was government-backed and was part of Microsoft’s £22 billion UK investment announcement in September 2025. The plan was 23,040 Nvidia GB300 GPUs in a 50MW facility, running from 2027. Microsoft reportedly ended its agreement with Nscale after reports that the site might not get enough grid power until the early to mid-2030s. Microsoft didn’t comment. Nscale says it is “exploring alternative power sources for the site.” The Register reports that demand for new grid connections surged from 41GW to 125GW in less than a year. Government backing is a promise. The connection date is a fact, and Microsoft can take its demand to any region that can plug in sooner.

Here’s the part of the week that doesn’t fit neatly. In the banks story, the promise hasn’t been tested yet. Goldman Sachs, JPMorgan and Morgan Stanley have made a combined $52 billion in equity trading revenue this year, nearly 80% of what the top five banks made. At Goldman, equities financing jumped 96% in the third quarter. The reassurance is that big balance sheets absorb losses. That’s true for losses that hit one client at a time. It’s much less true when most of the book is levered to a single theme. Code got its reckoning this week, and so did a privacy claim and a national infrastructure plan. Leverage on the AI trade hasn’t had one. The first signs of it are coming from lenders. HSBC CEO Georges Elhedery put it plainly: “One has to be selective.”

So the week settled one argument and opened another. Settled: when checking gets cheap, a reassuring sentence stops being worth much, and the edge moves to whoever can show what they’re claiming. Open: whether credit gets repriced the same way code just did. My bet is it does, and that it shows up in financing terms before it shows up in earnings calls.

Next week, watch two things: how fast maintainers turn findings into fixes and proofs, and who gets financed, and on what terms, at the margin desk.

A promise is a bet that nobody checks, and AI just made checking cheap.