The news, as reported
Microsoft reported this week that a China state-aligned group it tracks as Storm-0940 used an AI framework called ARTEX to automate attacks against more than 2,250 Microsoft 365 tenants — in some cases reaching privileged access in under three hours. The agents did the reconnaissance, the sign-in attempts, and the follow-through, at machine speed, with little human input.
CrowdStrike reported four previously unknown criminal groups doing something quieter. Between May and August, they talked commercial AI models out of their own safety rules — posing as authorized security testers until the guardrails gave way — and used the models to run intrusions.
And Axios reported that JPMorgan, Morgan Stanley, Bank of America, Wells Fargo, and Citi are now restricting outside AI agents from signing in to customer accounts. The agents were getting in the way customers let their own software in: with the customer's password. The banks' problem is simple to state. When an agent signs in, they cannot tell whether it is acting on the customer's instruction — or on its own.
What the banks actually did
Strip away the headlines and the banks did two things, in order.
First, they asked what their AI could see, and what it could touch.
Second, they drew a line around the answer.
Every restriction in that Axios report is the second step. The first step is the one that matters, and it is the step almost nobody outside a bank has taken. The banks have security departments whose whole job is asking that question before something goes wrong. They asked it late, under pressure, with the incidents already in the news. They still asked it.
The same question, at your scale
Your business uses AI differently than a bank. The shape of the problem is the same.
Think about what your tools hold right now. The client file you pasted in to get a faster summary. The contract you uploaded for a plain-English read. The inbox, the CRM, the drive — every login you shared so an assistant could "just handle it." Each of those was a decision. Most of them were made in a hurry, on a Tuesday, and never reviewed again.
None of that makes the tools bad. It makes them unexamined.
An attacker with an AI agent does not need your business to be big. CrowdStrike's groups did not pick targets by revenue; they picked them by exposure — poor hygiene, open doors, permissions nobody had looked at. Microsoft's agents moved in hours. The question is not whether your business is worth attacking. It is whether anyone has ever mapped what your own AI can reach.
What the audit checks
The Selenova audit is the banks' first step, built for a business with no security department. Four checks:
- The tools. Every AI tool and assistant connected to your business, listed in one place — including the ones you forgot you approved.
- What you've pasted. The files, figures, and client information sitting inside those tools, and what the tools kept.
- Where follow-up leaks. The shares, exports, and connected accounts where information leaves without anyone deciding it should.
- What stays yours. What should be removed, restricted, or rebuilt — in writing, in plain language, in order of risk.
You end with a map: what your AI can see, what it can touch, and the line you choose to draw around both.
The offer
The banks paid departments to ask this question after the fact. The audit asks it now, for the price of a single pilot engagement.
The first three audits are $750. When the three are done, the pilot price ends.
Know what your AI tools can see.
→ Book the audit: selenova.info/audit
