Have you ever given a spare house key to someone you barely know? Maybe a contractor you hired once to fix the garbage disposal. Most of the time nothing bad happens, but if that contractor’s truck gets broken into and the key ring disappears, you’ve handed a key to whoever stole it, and you won’t find out until something goes missing.
That’s more or less what happened to Vercel. According to the Sophos AI Security 2026 report, an employee signed up for Context.ai, a third-party AI productivity tool, using a corporate Google Workspace account and granted the app “Allow All” permissions. When Context.ai was later compromised, the attackers inherited those OAuth tokens and used them to work their way into Vercel’s internal environments. Vercel disclosed the breach in April.
The employee wasn’t trying to do anything wrong and just wanted a tool that made the job easier, which is why almost anyone tries a new AI app. Sophos is betting plenty of companies have the same exposure without knowing it. On October 13 it made Sophos AI Defense generally available, and the company says it discovers the AI tools employees and applications are using, including the ones security teams never approved, and enforces policy wherever those tools run.
Shadow AI Is Everywhere
A Gartner survey of 302 cybersecurity leaders found that 69% suspect or have evidence that employees are using unsanctioned AI tools. A Cloud Security Alliance survey (commissioned by Token Security) found that 82% of organizations discovered shadow AI agents in their environments over the past year.
None of that surprises me. We went through a version of this when smartphones started showing up at work. Employees didn’t wait for IT to write a policy; they just brought their own devices, and IT had to catch up. The difference, though, is that a phone mostly sat in someone’s pocket. An AI tool connected to corporate accounts can read email, summarize documents, and send data somewhere you’ll never see it again.
Kyle Falkenhagen, senior vice president of product management at Sophos, said in the announcement that most security teams can’t answer basic questions about which AI tools are running, what data is going into them, and whether it lines up with policy.
How Sophos AI Defense Works
AI Defense runs on Sophos Fusion, the AI-native defense system Sophos launched in July. Sophos says it pulls signals from controls customers already run on the device, across the network, at sign-in, and in the browser to map AI activity. It weighs identity, behavior, and the data involved to sort routine use from activity that needs a closer look. Sophos says real-time guardrails inspect prompts and uploads as they happen to help keep sensitive data from walking out the door, and that activity feeds into Sophos XDR and MDR like any other alert.
Building on controls that are already in place makes sense, because nobody needs the latest shiny object with yet another console to babysit. That cuts both ways, though. Sophos says AI Defense extends Sophos Workspace Protection, Endpoint, Firewall, and MDR, so if your endpoint or firewall comes from somebody else, there’s a good chance AI Defense has less to work with.
Attackers Use AI Too
In May, Sophos found an attacker it tracks as STAC6994 running roughly a dozen AI agents inside a customer’s network. The agents built and tested malware against Sophos, CrowdStrike, and Windows Defender and produced nearly 80 modules. The same operator later deployed ransomware and stole data. Sophos notes the tradecraft wasn’t new, but leveraging AI agents enabled attackers to turn weeks of trial and error into days.
The report also documents threat actors selling brokered API keys for Claude, ChatGPT, and Grok. Every unsanctioned AI tool holding an OAuth token or API key is another spare key floating around for those folks to grab, which is why the report says to start with an inventory.
Customers Are Starting to Ask
Anthony Bender, CTO at Empyreal Logistics, said in the announcement that his customers started asking how the company was using AI and how it was protecting it, and at first he didn’t have a good answer. He says AI Defense now lets the company show customers it’s doing what it says.
Shadow AI is an issue that every company is facing. Could you just ban AI and be done with it? You could try, but employees routed around bans on personal Dropbox and Gmail accounts years ago, and they’ll do the same with AI tools (probably faster). Sophos is pitching the other route, which is letting people use AI while security keeps an eye on it.
- Sophos AI Defense Goes Looking for Shadow AI - October 20, 2026
- How a Hacker Helped Mariana van Zeller Find Her Romance Scammer - October 16, 2026
- The Technology Tax You Pay Every Time You Buy Something ‘Smart’ - October 14, 2026
