I spent this week with three owners who asked the same question three different ways: what does AI actually cost me? Not one of them had opened the billing page in a month. That is where today's lead story starts. Follow the invoice, not the pitch, because the invoice has never once been wrong about you.
Google's threat intelligence group spent the last month watching dark web marketplaces sell access to models from Anthropic, Google and OpenAI at discounts of up to 97%. What is changing hands is not software. It is somebody's API key, lifted out of a repository or a support ticket and resold by the token. Techmeme carries the Financial Times report.
Demand for stolen access is high because AI capacity is expensive and scarce. Anthropic committed $11.6 billion over seven years to Akamai for computing capacity this month, per TechCrunch. Any market with that much demand and a payment instrument you can copy develops a black market, and this one now has a name: LLM-jacking.
So what: open your AI vendor's billing page this week and look for usage inside hours your team was asleep. Then rotate every key and set a hard monthly cap on each one. That is a twenty minute job, and every major vendor already ships the controls. If your team passes one key around inside a group chat, rotate it today, because the bill arrives in your name whether or not you were the one doing the spending. Nobody audits a key they did not know existed.
Do this: Put every AI key you pay for in one list, with a named owner and a cap.
PASTE THIS
List every API key on my account. For each give me: the key name, the date last used, this month's spend, whether a monthly budget limit is set, and whether an alert fires at half that limit. Output one table sorted by spend, highest first. Then repeat the table showing only the keys with no budget limit, and flag any key I cannot map to a named person on my team.
You'll get: A spend table most owners have never seen, plus the keys with nobody's name on them. Two of those are usually a tool someone tested in March and forgot. If the list comes back clean, you have proof your stack is capped.
Criminals are shipping fake HR desktop apps
Attackers spent this month impersonating HR and payroll providers with fake desktop apps, talking HR staff into granting remote access to their own machines. The tell is simple: none of those providers ships a desktop app. The Register has the attack chain. So what: tell whoever owns payroll before Monday that no HR vendor will ever ask them to install a desktop app, then ban new installs on that machine this quarter.
A court let the Pentagon keep Anthropic blacklisted
A federal appeals court upheld the Pentagon's ruling labelling Anthropic a supply chain risk, in a 2-1 decision. A US agency can declare your AI vendor a risk and win the first round. CNBC has the decision. Anthropic publishes an enterprise safeguards policy you can point at in your own review. So what: line up a second vendor for any AI function on a regulated workflow, before your next renewal date.
Microsoft is betting you want one AI app
Microsoft is folding its business AI tools into a single Copilot app, aiming at Anthropic's enterprise accounts. CNBC reported the strategy and The Verge covered the launch. So what: use the consolidation as your renewal lever this quarter. Ask one vendor to price every tool you buy separately as a single seat count, and hold it against what you pay today.
The six questions you send your AI vendors on Monday
Send this as plain email to every AI vendor you pay, and cc your own finance lead. Answers in writing, please. 1. Which of your models can take actions outside a sandbox? 2. What do you log for each agent action, how long do you keep it, and can I export it myself? 3. Who notifies me when an agent does something unintended, and inside how many hours? 4. Do you support per-key spend caps and alerts, or only account-level limits? 5. If a key of mine leaks, how fast can we revoke it together? 6. What model misbehavior did you disclose publicly in the last ninety days? Anything you get back in writing becomes the evidence pack your insurer and your auditor will eventually ask you for. The vendors who answer badly are the ones to move off, and the ones who answer well are worth a longer contract.
This week's question: we are twelve people and our AI spend went from nothing to $900 a month in six months. Cut it or accept it? Accept it, then make it visible. $900 a month is $75 a person, roughly what you already pay for one business tool nobody questions. The failure is never the amount, it is that nobody can state the number without opening a portal. Give AI spend one named owner and a cap per key, then review it beside payroll. Cost you cannot explain is a problem. Cost you can explain is a budget line.
How do you track what your team spends on AI tools?
A. One shared card, no breakdown
B. We see the total, not the detail
C. Per-tool budgets with hard caps
D. Honestly, nobody knows
Reply with one letter — results next issue.
Dark web sellers were offering stolen access to frontier AI models this week at a discount of up to what percentage?
Answer tomorrow — and no, we are not printing it here.
About 5 minutes to read. One thing you can use today every issue.
Every Friday I answer one reader's AI question in depth. Reply with yours — best one gets the full writeup.
Know an owner drowning in their inbox? Forward this — the 2-Minute Win is worth 2 minutes.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.
8735 Dunwoody Place STE R, Atlanta, GA 30350