Security
AI Is Becoming Dangerously Good at Hacking — and Governments Are Unprepared
New evidence from the UK AI Security Institute shows common AI models can be exploited by cybercriminals to automate sophisticated attacks at unprecedented scale.
By Patrick T ·
A new report from the UK AI Security Institute, published on May 25, 2026, has delivered one of the most sobering assessments yet of AI's role in cybersecurity. The report documents a rapid and largely unacknowledged shift: AI models that were designed for legitimate productivity applications are being repurposed by cybercriminals to automate sophisticated attacks at a scale and speed that existing defensive infrastructure cannot match.
The Capability Gap
The core finding of the report is what the authors call the "capability gap" — the growing disparity between AI-assisted offensive capabilities and AI-assisted defensive capabilities. While both sides of the cybersecurity equation are using AI, the report argues that offensive applications are advancing faster, for a simple reason: attacking is easier than defending. An attacker needs to find one vulnerability; a defender needs to protect against all of them.
The report documents specific cases in which commercially available AI models — including models from major providers that are widely used for legitimate purposes — were used to automate vulnerability discovery, generate novel malware variants, and craft highly personalized phishing campaigns. In controlled tests, AI-assisted attackers were able to compromise target systems in an average of 4.7 minutes, compared to 23 minutes for human attackers working without AI assistance.
The Democratization of Sophisticated Attacks
Perhaps the most alarming finding in the report is what the authors describe as the "democratization of sophisticated attacks." Historically, the most dangerous cyberattacks required significant technical expertise — the kind possessed by nation-state intelligence agencies or highly skilled criminal organizations. AI is changing that equation. The report documents cases in which individuals with minimal technical background were able to use AI tools to conduct attacks that would previously have required a team of experienced hackers.
This democratization has significant implications for the threat landscape. Nation-state actors remain the most sophisticated adversaries, but the pool of actors capable of conducting damaging attacks is expanding rapidly. The report estimates that the number of individuals capable of conducting significant cyberattacks has increased by approximately 300% over the past two years, driven almost entirely by the availability of AI tools.
We are entering an era in which the barrier to conducting a sophisticated cyberattack is no longer technical skill — it is intent.
UK AI Security Institute Report, May 2026
Critical Infrastructure at Risk
The report pays particular attention to critical infrastructure — power grids, water treatment facilities, financial systems, and healthcare networks. These systems are attractive targets for both criminal and nation-state actors, and they are often protected by legacy security infrastructure that was not designed to defend against AI-assisted attacks. The report identifies several specific vulnerabilities in critical infrastructure that AI tools have made significantly easier to exploit.
The healthcare sector receives special attention. Hospitals and medical networks have been frequent targets of ransomware attacks, and the report documents cases in which AI tools were used to identify and exploit vulnerabilities in medical device software. The consequences of successful attacks on healthcare infrastructure can be life-threatening, making this a particularly urgent area of concern.
Policy Recommendations
The report makes several specific policy recommendations. First, it calls for mandatory security testing of AI models before deployment, with particular attention to their potential for misuse in cybersecurity attacks. Second, it recommends the establishment of a rapid response capability that can identify and respond to novel AI-assisted attack patterns in near-real time. Third, it calls for international cooperation on AI cybersecurity standards, noting that cybersecurity threats do not respect national borders.
The report also addresses the AI industry directly, calling on model developers to implement more robust safeguards against misuse. This includes technical measures such as monitoring for patterns of use consistent with attack preparation, as well as policy measures such as enhanced know-your-customer requirements for access to the most capable models.
Industry Response
The major AI companies have responded to the report with a mixture of acknowledgment and defensiveness. Several companies noted that they already have extensive safeguards in place and that the report's findings reflect misuse of their systems rather than failures of their safety measures. Others committed to working with the UK AI Security Institute to address the specific vulnerabilities identified in the report.
The broader cybersecurity industry has been more receptive. Several major security firms have announced new AI-powered defensive tools in recent months, and the report's findings are likely to accelerate investment in this area. But the fundamental challenge remains: in a world where AI is making both attacks and defenses more powerful, the side with the structural advantage — the attacker — is likely to maintain that advantage unless significant policy and technical interventions are made.