Data Governance · 4 min read
31% of UK Organisations Hit by AI Incidents in Two Years
New AI Score data shows 31% of UK organisations suffered AI incidents in two years. Only 24% hold a full inventory of their AI tools. Governance gaps are widening faster than controls.
Key takeaways
- 31% of UK organisations experienced an AI incident in the past two years, per AI Score's survey
- 50% of affected organisations report unapproved shadow AI tools; 44% entered confidential data into AI applications
- Only 24% of businesses have a complete inventory of their AI and agentic tools
- 1% of phishing attempts now contain hidden prompt injection, per Bridewell; Microsoft's protection blocked all detected cases
- NCSC received 1,831 attack reports on UK organisations between September 2025 and August 2026, up from 1,727

Thirty-one percent of UK organisations have experienced an AI incident in the past two years. That figure comes from a new survey by London-based AI Score, and the details behind it are worse. Half of the affected organisations report unapproved "shadow" AI tools in use, and 44% say confidential or sensitive data was wrongly entered into AI applications. Regulatory breaches and biased or discriminatory output each account for another 21%. Unauthorised actions by AI agents account for 13%.
The pattern repeats across the survey. Fifty-nine percent of organisations claim they govern AI well. Yet only 24% hold a complete inventory of the AI and agentic tools they run, and only 35% have full visibility into what those tools can access. The gap between self-assessment and evidence is the story.
How bad is the governance gap?
It carries a commercial cost. Fifty-six percent of respondents say external parties increasingly request evidence of AI governance before doing business. Seventy-eight percent report rising investment in governance, and 84% now treat robust AI governance as a competitive advantage. Buyers are asking for proof. Most sellers cannot produce it.
The AI Score findings align with earlier work on shadow AI. A September Deloitte study found UK workers spend £958 million out of their own pockets on large language model tools (AI systems that generate text from prompts). Only 51% of AI-adopting employees use tools provided by their employer. The other half is invisible to security teams.
Alex Harland, CEO of AI Score, argues visibility is the foundation. "Businesses need to understand which and where AI systems and services are being used; what data is being accessed or processed; who is doing what and, increasingly, what actions are AI agents taking," he said. "That gives the evidence to decide where safeguards are needed, where action is required, or where progress is being made."
Harland recommends continuous monitoring rather than a one-off audit, since tools, permissions and use cases all evolve. He also urges firms to appoint a dedicated "AI manager" — not necessarily an executive, but a role with board-level backing and clear ownership of how the organisation uses AI. "Every organisation using AI should have someone clearly accountable for its governance, with the authority and resources to make that accountability effective," he said.
Is AI also fueling the external threat?
Yes, according to the incident data. The UK's National Cyber Security Centre received 1,831 reports of attacks on UK organisations between September 2025 and August 2026, up from 1,727 the previous year. Black October — the wave of industrialised hacking marked by PHP bugs and compromised supermarket fridges — fell outside that review period, and suggests 2027 may go higher.
Chris Anley, chief scientist at NCC Group, frames the problem as two-sided. "Businesses must contend with both the risks created by their own adoption of AI and the possibility that attackers will use the same advances to target them more effectively," he warned. He describes AI as a "significant" shift in cybersecurity because it expands the attack surface at both ends: defenders deploy models and agents with novel vulnerabilities, while attackers use frontier models to analyse codebases and find software flaws at scale.
"The concern is that the gap between bad actors discovering a vulnerability and exploiting it could narrow, putting existing patching and incident-response processes under greater pressure," Anley said. "AI does also give defenders more effective tools for detecting and responding to threats quickly, but this advantage is not automatic."
Martin Riley, CTO at UK cybersecurity firm Bridewell, sees the same pattern from the trenches. Businesses have deployed AI "faster than they have understood their dependency on it," he said. "Assistants are being connected to email, documents and operational processes because that is the quickest route to value. Security teams are then left working out what those tools can reach and what they can do."
Bridewell has already detected exploits aimed at AI assistants specifically. One percent of phishing attempts now contain hidden prompt injection (malicious instructions written for the AI reading the email, not the human). Ten percent of those attempts targeted critical national infrastructure clients; the rest targeted government bodies. Microsoft's prompt-injection protection blocked all of them. "The volume is low," Riley said, "but the signal matters."
What should data and security teams do?
The recommendations converge on one word: inventory.
- Map all AI assets. Cover embedded features, employee tools and autonomous agents. Record who owns each, what data it uses and which processes depend on it.
- Treat agents as identities. Give each agent defined permissions for specific tasks, exactly as you would an employee.
- Audit your attack surface honestly. Riley asks: "Are you releasing public-facing apps that need more DevSecOps attention? Are you spread across cloud environments with uneven skills and varying maturity?"
- Stress-test patching and response. Anley advises regular reviews of vulnerability management, prioritising critical exposures and ensuring response keeps pace with faster attackers.
Anley closes with a caution against tool-washing. "AI can strengthen threat detection, accelerate investigations and help security teams prioritise vulnerabilities, but experienced professionals remain essential for validating findings, interpreting risk and making informed decisions," he said. Analysts should read that literally. The model does not replace the judgement.
By Rebecca Stone
Google News: Data Governance (Source)
More from Rebecca Stone
Show full bio
Senior reporter covering regulation, policy and their impact on Analytics Digest is a professional daily at Analytics Digest.
12 articles