One-in-Five Breaches Involve AI, A Shift in Cyber Risk
IBM's 2026 Cost of a Data Breach report shows a striking trend: 22% of UK businesses faced an AI-related security breach in the past year. This pattern is not just seen in the UK.
Globally, AI-driven attacks now account for a significant share of incidents, and their cost is rising fast. Today, the average price tag for an AI-linked breach has reached $6 million.
These numbers signal a change in cyber risk. Attackers don't just go after the AI systems; they also use AI to make their attacks faster and more effective. Mark Hughes from IBM's Cybersecurity Services puts it plainly: “AI has dramatically lowered the barrier for cybercriminals.
Attackers can now execute attacks in minutes rather than days with advanced frontier models.” The speed is new. It changes everything.
AI-Enabled Threats: Deepfakes, Malware, and Shadow Agents
Among all forms of AI-powered attack, deepfake impersonation stands out as the leading tactic, with 45% of organisations surveyed by IBM reporting such incidents. Deepfakes are not alone, however; research on LLMs and generative AI in cybersecurity highlights that large language models now drive automated phishing attempts, malware creation, wide-reaching social engineering campaigns, and new ways to identify software vulnerabilities.
The threat goes deeper once inside company networks. The Cloud Security Alliance reports 82% of enterprises have discovered unknown or “shadow” AI agents running within their IT environments, often with no formal oversight at all.
These shadow agents can linger long after they've served any useful purpose. Some still hold credentials and permissions that put companies at risk, a hidden “retirement debt” that may result in major breaches later.
Visibility Gaps and Lifecycle Risks
Many organisations believe they know what's running on their networks. In reality, visibility often falls short. Although 68% of CSA survey respondents say they feel confident about tracking their AI agents, most also host unknown or unmanaged agents without realising it.
Most shadow deployments show up as internal automation scripts or custom LLM-powered utilities. Some slip into SaaS products or workflows built by developers themselves.
Agent lifecycles present another weak spot. Only 21% of surveyed enterprises have formal decommissioning processes for their AI agents in place. As tools gain autonomy, and with greater permissions, the absence of structured end-of-life planning leaves organisations exposed to data leaks, disrupted operations or steep financial losses.
The Governance Gap: Controls Lag Behind Adoption
The adoption rate for new AI tools outpaces the ability to secure them effectively. In its 2026 SANS Institute Survey, use of AI in cybersecurity jumped globally from 50% to 78% among professionals within a year, yet over half admit they still lack audit frameworks for managing enterprise-wide AI systems.
This mismatch creates strain for security teams who inherit broad governance responsibilities without adequate technology support behind them. On the ground, this is obvious: 63% reported serious weaknesses when detecting or responding to threats involving AI tools, a number higher than last year's.
Attackers Innovate Faster Than Defenders
Certain tactics evade defences because attackers don't use AI once. They employ it throughout every stage of their operations. Deepfakes and synthetic content enable highly convincing impersonations for fraud or executive-targeted phishing campaigns (often called business email compromise).
Meanwhile, adversarial techniques target machine learning models directly and trick them into misclassifying data or leaking sensitive information.
A review of emerging risks highlights four main categories:
- Deepfakes & Synthetic Media defeat human verification, distort information environments or extort targets with manufactured evidence.
- Adversarial Attacks manipulate inputs specifically to evade detection systems or disrupt model results.
- Automated Malware generated by LLMs evolves too fast for signature-based defences to keep up.
- AI-Powered Social Engineering enables mass phishing campaigns personalized through public data scraping and rapid profiling methods.
No Silver Bullet, Defensive Measures Are Still Catching Up
NIST's analysis from May 2026 shows agreement among industry leaders: traditional cybersecurity practices like least privilege remain valid but require adaptation as agent autonomy increases further.
You can't just check permissions at deployment then stop; controls need to follow agents through every use phase right up until decommissioning as well. There isn't one tool that fixes everything immediately, instead, policy enforcement at critical moments (such as requiring human signoff when an agent acts outside its role) proves most effective.
Nearly four out of five CSA survey participants made context-aware controls their top priority for the coming two years. The SANS Institute backs this approach: skilled people remain central to defence today.
Nearly half prioritized behavioural detection strategies; others picked user awareness training or direct analyst supervision instead. Those human-led methods shift when attackers do. No single tool adapts quickly enough alone.
Investments Are Rising, But Outcomes Are Uneven
Breach costs stay high but drop measurably when firms invest in incident response plans, stronger data protection initiatives or focused employee training after an attack hits. IBM reports nearly two-thirds of UK businesses expect to boost cybersecurity spending post-AI breach; worldwide that figure rises to 85%, according to follow-up research by the Ponemon Institute cited in IBM's report.
The Path Forward: From Discovery to Operational Defence
No company surveyed by CSA escaped serious business consequences after an incident caused by an unmanaged agent, be it leaked data or lost funds.
Progress depends less on finding new risks than on governing agent behaviour at scale: managing permissions proactively, decommissioning unused agents rapidly, and enforcing live rules wherever possible. The focus now is on continuous checking.
Deploying new tools helps, but embedding monitoring lets companies see how well models stand up against actual attack attempts over time. Building technical defences alone won't work. Workforce training matters just as much.
Those who prepare analysts with modern threat knowledge, and deep familiarity with current-generation AI, won't just block attackers more effectively but also increase returns from investments in AI & automation projects.
Cover photo by cottonbro studio on Pexels
