Shadow AI Creeps Past Security Teams

At many organizations, the security team is likely looking at an incomplete picture. That’s the warning from the UK’s National Cyber Security Centre (NCSC), which on September 7 published a blog post detailing the risks of “shadow AI” — the use of artificial intelligence tools that haven’t been vetted or approved by an employer.

The NCSC’s guidance arrives amid mounting evidence that unapproved AI has quietly become common practice in British workplaces. The agency cited Microsoft research showing that 71% of UK employees have used AI tools their employer never sanctioned — a signal that staff are moving faster than corporate oversight.

A Form of Shadow IT

Shadow AI sits squarely within the older concept of shadow IT: any technology deployed inside an organization without going through official channels. The difference now is the speed and scale at which generative AI services spread, often before enterprise policies can adapt.

The NCSC cautioned that when employees feed company or customer data into these unapproved tools, they may be handing over information that is then stored, retained, or used to improve the provider’s service. That effectively strips the organization of visibility and control over its own data — enabling potential breaches, intellectual property loss, and failures to meet regulatory requirements.

Why Employees Turn to Unapproved Tools

The agency’s post identified a key driver: existing cybersecurity policies sometimes fail to satisfy real business needs. In those gaps, staff improvise with whatever AI tool is available, adopting new services before their employer has had a chance to assess them.

The NCSC believes this pattern is likely to persist. New AI services are emerging faster than most organizations can evaluate them and offer sanctioned alternatives. The result, the NCSC said, is a persistent blind spot for IT security teams.

AI Agents Add New Danger

The warning extended beyond chat-based tools to a growing class of software: AI agents. These are autonomous systems that can act on behalf of a user, and the NCSC said they carry their own critical vulnerabilities. An attacker who manages to exploit a weak agent could gain the same data, services, and privileges the agent legitimately holds.

The blog post went further, asserting that attackers are highly likely to target agents with looser guardrails to exploit vulnerabilities or misconfigurations elsewhere in the corporate IT environment. In effect, an unmanaged agent can become a doorway into the broader network.

Reduction, Not Elimination

The NCSC’s advice is pragmatic: organizations should focus on reducing shadow AI rather than trying to eliminate it altogether. That mirrors how security teams have learned to handle shadow IT more broadly — you can’t block every path, so you manage the risk.

The agency recommended fostering a positive cybersecurity culture, one in which employees feel comfortable discussing security issues openly. The goal is to move the conversation about AI tool adoption out of the shadows and into a structured process.

“Organizations can’t hope to block connections to all possible AI tools, so they need to develop a positive cybersecurity culture with open dialogue about the tools staff might wish to use and to set clear guardrails around what secure use of AI looks like,” Chismon said.

— David Chismon, NCSC CTO for architecture

Chismon also warned earlier in the post that “IT security teams should not assume they are seeing the full picture,” underscoring the visibility gap that shadow AI creates.

International Agentic AI Guidance

The NCSC pointed to external guidance on the careful adoption of agentic AI services, which it published with international partners. This suggests the risks of autonomous AI are seen as a shared global concern, not limited to UK organizations.

For security teams, the message is clear: the rollout of AI in the enterprise is happening whether or not it’s approved — and that makes it a governance problem as much as a technical one.

Why It Matters

For CISOs and security managers, the NCSC’s warning suggests that the perimeter has effectively moved. Employees are already using AI services their employers don’t know about, and that gap between sanctioned tools and real-world behavior is where data leaks and compliance failures are likely to begin.

The 71% figure from Microsoft research offers a measure of how entrenched this behavior has become. If most UK employees have already crossed that line, then the question is not whether shadow AI will cause an incident, but when — and whether the organization will have the visibility to detect it in time.