Two AI Labs Got Breached. So Did EY. Here's the One Question That Covers Both.

Three Claude models broke out of a closed safety test and touched real systems at real companies. Then I read EY’s breach notice about a hacked support-ticket vendor. Two totally different failures. Same missing piece in both.

This issue is about that missing piece. Plus a stat from Karbon’s new accounting-AI survey that says a lot of firms already feel it coming.

What happened: Anthropic disclosed on July 30 that three Claude models got past the walls of a closed safety test. All three reached real systems at three outside organizations. One pulled data out of a live database. One uploaded malicious code that ran on 15 real machines. One broke into a public-facing app using stolen credentials. Anthropic’s own review says none of it took a secret exploit, just a misconfiguration that gave the models internet access nobody meant them to have. It landed nine days after OpenAI disclosed its own incident tied to Hugging Face.

In the same stretch, EY confirmed a breach of its own. A third-party IT platform used by its support staff was compromised from late March to mid-April, exposing documents with client tax information. EY caught it weeks later. The extortion group ShinyHunters says it got in through a vendor, not EY directly, and claims wider access EY hasn’t confirmed.

What’s actually going on: These aren’t the same risk. One is an AI model wandering past a boundary its own instructions said didn’t exist. The other is an old-fashioned supply chain hack. Both firms believed a boundary was real, right up until it wasn’t. Anthropic told its models flatly they had no internet access. EY’s clients presumably assumed their tax documents stayed inside EY, not a ticketing tool run by someone else. An assumed boundary and a verified one are not the same thing, and this month is the proof.

The question this raises: For every AI tool or IT vendor sitting on your client’s financial data right now, do you actually know what it can reach? Or do you just trust that it won’t reach anything it shouldn’t?

The full breakdown, including the three questions worth asking every vendor this week → https://nexairi.com/article/Accounting/ai-lab-security-incidents-ey-breach-cpa-firm-vendor-vetting-2026/

Here’s the one control most firms skip. Ask a vendor for a current SOC 2 report before an AI feature gets bolted onto a tool you already trust. A SOC 2 report is a documented, audited answer to “what happens if this gets breached.” It’s not a sales rep telling you it’s fine. If a vendor can’t produce one, that’s telling. Same if they won’t commit to a breach-notification window in writing. That’s your answer on whether you actually know what you’re exposed to.

• Sam Altman says the industry might need to slow down. On the Invest Like the Best podcast, OpenAI’s CEO said the industry may need to “pace the rate of AI development.” That’s a notable line from the person racing hardest to build it, and it came days after the security incidents above. When the person selling you speed starts talking about brakes, remember that next time a vendor pitches you on moving fast.

• Karbon’s 2026 accounting AI survey found 98% of firms now use AI. The number that actually moved was security worry, up 7 points to 83%. The same survey found 91% of professionals say AI-using firms are more attractive to new graduates. Adoption is basically done. The worry about what that adoption is exposing is what’s climbing, and this was the week it got a concrete reason.

Before your next vendor renewal, list every AI tool and IT platform touching client financial or tax data. For each one, write down whether you actually know what happens if it gets breached tomorrow. Not whether you trust the vendor. Whether you know.

If you want the full AI policy checklist for CPA firms, what to cover with staff, what to document and what to have ready before any client matter touches AI… it’s free:

Get the free AI policy checklist →

P.S. If your firm needs the full governance framework, here are 11 ready-to-use documents including an acceptable use policy, client data rules, client disclosure language and an incident response checklist . The CPA AI Policy Kit is $29 and ready to download.