A few weeks ago, Summer Yue — Meta’s Director of AI Alignment, the person literally responsible for making sure AI does what it’s told — gave an AI assistant called OpenClaw access to her email. She asked it to review her inbox and suggest what to archive or delete. Crucially, she told it not to take any action until she gave the go-ahead. It mass-deleted her emails anyway, writes David Shepherd, SVP of EMEA at Ivanti, pictured.
She told it to stop. It kept going. She had to sprint to her computer and manually kill the processes to shut it down.
Really let that sink in for a moment.
If someone whose job title includes “AI Alignment” can’t stop an AI assistant from steamrolling her own instructions, it’s fair to ask: how ready are the rest of us?
AI assistants — OpenClaw, Copilot, ChatGPT and the rest — are genuinely useful. That’s why adoption has been so rapid. Ivanti’s research found that 42% of office workers now use generative AI tools at work, up 16 points from the prior year. People are managing their calendars, summarising documents, drafting emails and generally getting through their to-do lists faster. Good.
But (and it’s a significant but) much of this is happening without anyone in IT or security knowing about it. Forty-six per cent of office workers say some or all of the AI tools they use are not employer-provided. Among gen AI users, nearly a third are actively hiding their usage from their employers — some because they like the competitive edge, others because they worry about their jobs.
So, we’ve got a situation where employees are feeding workplace data into tools that haven’t been vetted, aren’t being monitored and may not meet basic security or data protection requirements. Nobody is doing this maliciously. They’re trying to work smarter. But the security implications are real, and they’re growing.
Governance isn’t keeping pace
Ivanti’s 2026 State of Cybersecurity Report found that less than two-thirds of UK security professionals employ formal guardrails to oversee the deployment and operation of AI systems and agents. I think that’s a pretty telling number. It means a substantial proportion of organisations have no structured oversight for how AI tools are being introduced and used — even as those tools gain deeper access to email, documents, calendars and internal systems.
And the risks aren’t theoretical. The OpenClaw episode with Summer Yue showed what happens when an AI assistant oversteps its instructions in plain sight, with the user watching in real time. Now consider the shadow AI scenario: an employee pastes a confidential client brief into an unsanctioned chatbot, or feeds proprietary data into a free tool to speed up a report. No one in the business even knows it happened.
Locking everything down won’t work
The tempting response, perhaps, is to ban unsanctioned AI tools outright. It sounds decisive. In practice, it tends to drive the behaviour further underground. Employees are already finding workarounds — and blanket restrictions can actually make the secrecy worse.
What does work is practical training that goes beyond a compliance tick-box exercise. Employees need to understand why sharing certain data with AI tools creates risk — not in abstract terms, but in specific, relatable ones. What happens when a free chatbot ingests your client’s financial data? Who owns that data now? What are the regulatory implications?
Ivanti’s research found that nearly half of office workers say they’re left to teach themselves how to use new technology. When the technology in question can access and process sensitive business information, self-taught isn’t good enough.
Leadership more than anything
It would be easy (and unfair) to blame employees for all of this. They’re using AI because it helps them do their work. Many are keeping quiet about it because their employers haven’t created an environment where it’s safe to be open.
That needs to change from the top. Clear policies on which tools are approved and which aren’t. Transparent guidance on what data is and isn’t appropriate to share with AI platforms. Regular updates as the technology and its risks evolve. And above all, a culture where someone can say “I’ve been using this AI tool — is it secure?” without worrying they’ll face judgment or backlash.
Security, HR, legal, and IT teams need to work on this together. If a policy only lives in a security handbook that nobody reads, it’s not a policy at all. Guardrails need to be practical, communicated clearly and built into the way people really work.
AI assistants are going to keep getting more capable. That’s broadly a positive development. But every new capability comes with new questions about data access, permissions and oversight. If an organization treats AI governance as an afterthought, that organization is begging for exposure. The Summer Yue incident is a reminder that even experienced professionals can get caught off guard when the guardrails aren’t there.
