Anthropic's Enzyme Discovery Vanished on the Second Run

So, What Actually Happened?

Friday, and I keep turning over three things from Thursday that I first read as unrelated. We scanned 190,000 articles this week so you don't have to, and these three kept sitting next to each other. An AI model announced a new CRISPR-like enzyme system, and then ten reruns could not find it again. Maine's attorney general pulled its public breach database offline after discovering the notices were fabricated. In Australia, an agent cleared Medicare's portal controls without breaking anything. Then Gartner put next year's AI spend up 49.5%, a number set before any of that was known.

The Bottom Line: Claims are arriving faster than anyone can check them, and next year's budget already assumes they are true.

Advertisement

The Tracks That Matter

1. An AI Found A New Enzyme System, Then Lost It Again

Anthropic said its model turned up a CRISPR-like enzyme family no human had described, which would be the first serious claim of an AI making an original biological discovery. Then someone ran it back. Ten reruns of the same setup could not reproduce the finding, and nobody yet knows whether the first run found something real or assembled something plausible. Both stay possible, and that is the uncomfortable part, because the discovery was announced on a single run. Any lab would file one unreproduced result as a hypothesis. The press cycle filed it as a discovery, and your board read the press cycle.

Here's what works: When a vendor demos an AI result, ask how many times they ran it and what the other runs produced.

2. An Agent Walked Straight Through Medicare's Front Door

An OpenAI agent reached records inside Australia's Medicare portal that it should never have touched. Nothing was hacked in the old sense. The agent passed the portal's identity checks because those checks test whether someone knows the answers, not whether someone is a person, and something holding the answers passes every single time. Darktrace shipped rogue-agent detection for enterprises the same morning, which tells you where the vendors think this goes. And this was a national health system with a real identity layer, working exactly as designed.

Here's what works: Take the verification step you trust most and test it against an agent this week, not against a person. Knowledge-based checks fall first.

3. Gartner Puts Next Year's AI Bill Up Another 49.5%

Gartner expects AI spending to jump 49.5% in 2026, which is less a forecast about enthusiasm than a readout of contracts already signed. Underneath it the money is turning physical: data centres are now financed like infrastructure assets rather than IT projects, with the long tenors and covenants that come with that. Once a build is financed over fifteen years, nobody unwinds it because a pilot disappointed. That is the real lock-in, and it has nothing to do with model quality. Half again as much spend, committed before this week's results were checked, is a very large decision made on trust.

Here's what works: Split next year's AI budget into what is contractually committed and what is still yours to cancel. Most teams have never drawn that line.

Quick hits:

• Maine published fake breaches, then took its own database down. The state attorney general removed hoax notices for VRChat and Discord and temporarily disabled its public breach database, which is worth remembering the next time a vendor questionnaire cites a public breach record as evidence.
• Agent permissions are now something you acquire, not build. Omada bought EmpowerID to police AI agents in real time, which puts identity governance for non-human accounts on the M&A sheet rather than the roadmap.
• One robotics startup shipped its evidence alongside its claim. Light Origins released Light-O1 with prediction-error curves measured across different robot bodies, which was rarer this week than a model launch.

Advertisement

Signal vs. Noise

🟢 Signal: the governance function. Governance gained more real hold on everything else this week than any other idea we track, while being written about less. The evidence is operational rather than rhetorical: an identity vendor buying agent controls, a privacy regulator standing up an audits division, a security vendor shipping rogue-agent detection. The people who check the work are quietly becoming the people who approve the spend.

🔴 Noise: ”data security” as a category. It pulled heavy volume and lost ground where it counts. Almost nothing that went wrong this week was an encryption problem or a perimeter problem. Medicare was an access failure. Maine was a verification failure. The category name no longer matches the incidents being filed under it.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A policy group argued that nobody can measure AI adoption, a security team published the first real vulnerability counts for the protocol agents use to reach your tools, and a Finnish university put out a usage report on its own assistant, all inside one morning.

Three desks, three stories that never meet. Policy press covers the case for a Google Trends for AI. Security press covers what the MCP vulnerability data reveals. Nobody covers a university documenting its own assistant, because why would they. Read on the same morning they say one thing: the instruments for measuring AI are being built right now, by the people who already deployed it, years after they deployed it. That is why every adoption statistic you have been shown this year has no denominator underneath it. Not because vendors are lying, but because there is nothing yet to count against. The organisations building their own measurement are doing it because nothing external exists to buy.

What changes Monday: next time a vendor tells you what percentage of teams use their product daily, ask what sits at the bottom of that fraction. The answer teaches you more than the number.

Advertisement

By The Numbers

• AI spending will jump 49.5% in 2026 — Gartner's figure, and it was set before anyone tried to check this week's results.
• ChatGPT referral traffic more than doubled in 2026 — if your acquisition model still assumes search sends the visitors, it now describes the smaller half of the internet.
• Los Angeles County settled with Walmart and Sam's Club for $908,817 — tracking pixels that leaked health-related purchases, and the going price of a marketing tag nobody reviewed.
• SPRIND and NADI put €40 million into AI-native chip design — Europe funding the design step instead of trying to buy its way into fabs.
• More than 1,500 developer applications arrived in under a week — for connectors into a consumer agent that is getting its own email address and a payment rail.
• See what's rising across AI and data this week →

Deep Dive: Who Checks The Work

I interviewed at McKinsey once for a data modeler role. They took my SQL Server and ETL skills for granted after a brief explanation, then put an Excel sheet in front of me and asked how I look at data. I talked them through filtering, pivot tables, profiling. Two hours, a rate discussion at the end, hired. I never had an interview that good again, and they never once asked which tools I owned.

A result you cannot repeat is a hypothesis

One run produced an enzyme family. Ten runs did not. In any lab that is the normal, boring middle of science and nobody calls it a finding yet. What changed this week is the distance between the first run and the announcement. It used to be months of replication. Now it is hours.

A control you never retested is decoration

Medicare's portal was not weak. It asked the questions it was built to ask and something answered them correctly. Knowledge-based verification assumes whatever sits on the other end is slow, forgetful and human. None of those three assumptions hold anymore, and nobody went back to retest the control when they stopped holding.

The money does not wait for the receipt

Spending up by half again, data centres financed on fifteen-year horizons, agent permissions changing hands in acquisitions. All committed while the verification layer is still being invented. That is not recklessness, it is sequencing, and the sequence puts the bill before the proof.

What Actually Works

1. Ask for the rerun count: for any AI result a vendor shows you, ask how many times it ran and what the other runs produced. One run is a hypothesis wearing a press release.
2. Test a control against an agent: pick the verification step you trust most and point an agent at it. Knowledge-based checks are the first to fall, and you want to find that out yourself.
3. Demand the denominator: every adoption percentage has a bottom half. If nobody can produce it, the number is marketing with a decimal point.
4. Name a human per automated claim: every agent and every AI-generated finding gets one person who owns whether it is true. Not a team. A name.

Anyone can land one track by accident. A set is the part you have to do again.

What's Coming

Financial Supervisors Start Naming AI Directly

A new spotlight on AI in financial services reads more like the draft of a supervisory expectation than a market update. When law firms publish sector-specific AI guidance for banks, the questions in the next examination cycle are already written. Expect model documentation requests well before you expect legislation.

Governed Agents Become Something You Buy

Valtech launched an agent factory with governance built in, which is what a category looks like shortly before analysts draw a square around it. The pitch has moved from what agents can do to who signed off on them. If your agent programme is still a platform decision, it becomes an audit decision by spring.

Open Weights Arrive With Strings

Beijing's BRICS pitch offers open weights inside closed loops: the model is free, the surrounding infrastructure is not. Procurement teams evaluating cheap open models next year will discover the licence was never the expensive part. Read the deployment terms, not the weights licence.

For Your Team

Monday's meeting prompt: ”Name one AI result we acted on this quarter. How many times was it reproduced before we acted, and who personally checked it?”

Share-worthy stat: An AI model announced a new CRISPR-like enzyme system, and ten reruns of the same setup could not find it again. Both facts landed in the same week. The first one travelled far further than the second, and that gap is the whole problem with how AI results reach a board.

Go deeper: Track where AI claims and AI money are actually moving →

The Track of the Day

”The creation of CalPrivacy's Audits Division is designed to look under the hood of businesses' privacy and cybersecurity practices before problems arise.”
California Privacy Protection Agency, on its new audits division

A sound engineer checks the room before the doors open, not after the complaints come in. Almost everyone building with AI this year is working in the other order.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: September 25, 2026 | Curated by Yves Mulkers @ Ins7ghts