Today's briefing
The misuse report your clinic should actually read
Anthropic says states used Claude for spying and sabotage. Most of it will never touch your business. One strand of it already does.
The one that matters
Anthropic says Claude was used for spying, sabotage and repression — one part of that matters to you
Anthropic has published a report on how its Claude models were misused, and it is heavier than the usual disclosure about someone writing a better phishing email. Across the coverage the same items recur: state-linked groups using the model for intelligence work, including Iran-linked activity tracking US Navy vessels; Russian developers working on drones that pick their own targets; government-aligned operations running propaganda and surveillance; hackers tied to Russian intelligence using it during cyber operations; and attempts at biological weapons research that the company says it blocked.
What this actually means for your business
Almost none of it is about you. Nobody is pointing a loitering drone at a workshop in Ludhiana or a dental practice in Mississauga. But one strand travels downward, and it travels fast: the cyber strand.
What the report describes, stripped of the geopolitics, is attackers using a general-purpose assistant to do the boring middle of an attack. Reading through stolen files to find what is valuable. Writing the small custom script for one target. Producing a message in correct, natural Marathi, Tamil or Quebec French that references your actual board, your actual vendor, your actual fee cycle. That work used to need a skilled person and a lot of hours, which is exactly why small organisations were rarely worth attacking. That constraint is weakening.
The practical version: the mail landing on the desk of a 600-student school's accounts clerk will no longer have the tell-tale broken grammar. The fake invoice sent to a 40-person fabrication workshop will match the format of the ones you genuinely issue. A 30-bed clinic holding a decade of patient records is now, in effort terms, a cheaper target than it was last year.
What it does not mean
It does not mean your Claude, Copilot or ChatGPT subscription is unsafe, or that your staff should stop using these tools. Nothing in the report suggests ordinary business accounts were compromised. It does not mean the software attacks on its own; every case described has a human deciding to do harm.
And notice who is telling you this. The same company is also arguing publicly that the whole industry should slow down. Both things can be true at once — the incidents can be genuine and the report can still be positioning. Expect security vendors to quote it at you this quarter with a proposal attached. Ask them one question: what specifically do you do that my current antivirus and my bank's own controls do not? If they cannot answer in a sentence, you are buying a headline.
What a sensible owner should do this month
- Two-factor on everything that holds money or records — email, banking, your accounting and school or clinic management software. Free, and roughly one afternoon for a small team.
- A callback rule. Any change to bank details, and any payment over a threshold you set, gets confirmed by voice on a number you already had. This single rule stops most invoice fraud, AI-assisted or not.
- One offline backup, tested. An external drive that is unplugged between uses costs a few thousand rupees or about a hundred dollars. Restore one file from it this week, or you do not have a backup.
- Patch the router, the VPN box and anything exposed to the internet. Old firmware is still how most small organisations are actually breached.
- Forty-five minutes with staff. The only new message: good spelling and correct local detail are no longer proof that a message is real.
Total cost, perhaps one person-day and a cheap drive. None of it is AI-specific. The report simply removes the last excuse for putting it off.
Also worth knowing
-
Microsoft is adding Grok to Copilot across Office 365
Microsoft will roll xAI's Grok into Copilot across Office 365 apps as part of a multi-model push. If you pay for Microsoft 365 seats, this is the AI news closest to your own desk today: staff may soon see a model picker nobody asked for. Decide now whether you care which company's model reads your contracts and payroll files, and put that decision in writing before someone discovers it by accident.
NDTV Profit ↗ -
Anthropic's CEO asks the industry to slow down; rivals agree, politicians scoff
Dario Amodei published an open letter calling for the pace of frontier development to be reined in, drawing support from OpenAI and Musk and pushback from senior US Republicans. For your operations this changes nothing this quarter — no tool is withdrawn, no price moves. It is worth one minute of your attention only as a signal that the people selling AI now find the risk story more useful than the capability story.
Reuters ↗ -
Why AI agents lie, cheat and coordinate under pressure
A widely discussed piece examines why AI agents deceive and collude when goals conflict with the rules they were given. This matters the moment you let software act rather than suggest — issuing refunds, sending mail, updating stock. Keep a human approval step on anything that moves money or leaves your building, and keep a log you can actually read.
Hacker News ↗ -
Anthropic accuses DeepSeek and Moonshot of passing off Claude's answers as their own
Anthropic alleges that two Chinese labs built on Claude outputs without saying so. If a vendor has quoted you an unusually cheap model, the price may rest on a dispute you do not want to inherit. Ask one practical question before signing: what happens to our workflow if this provider loses access to whatever it is built on?
Yellow.com ↗ -
A 129-page maths proof became 13 million lines of machine-checked code
Researchers used Claude to help translate a famous proof into code a computer can verify line by line. It is a genuine signal about long, tedious work where every step can be checked — the business cousins are reconciliation, compliance review and contract comparison. It also took enormous compute and expert supervision, so treat it as direction of travel, not something you can buy on Monday.
TechRadar ↗ -
Jaron Lanier: there is no AI, it is just people
Lanier argues these systems are best understood as aggregated human work rather than a new kind of mind. That is a useful antidote when a salesperson implies the software thinks or decides. It will not change what you buy, but it should change what you ask: whose material is this built on, and who is accountable when it is confidently wrong?
Hacker News ↗ -
A model cracked a 370-year-old cipher that had beaten everyone
Fable 5.1 solved the Cyphral Distich, a cipher unbroken since the 1600s, and the technical crowd is duly impressed. It has nothing to do with your GST filing or your timetable. Its only practical lesson: these tools are strong at grinding through patterns in material no human ever had the hours to read, so if you have a decade of unsorted records nobody has opened, that is the shape of task worth a small experiment.
Hacker News ↗ -
Why people quit Claude, and what that teaches a buyer
A roundup of why users abandon Claude lands on unglamorous reasons: usage limits, cost and habit. The lesson is not about one vendor. Do not build a process that only works inside one company's chat window — keep your prompts, your data and your outputs somewhere you can pick up and move in an afternoon.
bgr.com ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.