By 2026, documented cases had begun to show how the same incident could involve AI-driven tools on both the offensive and defensive sides of a cyberattack, with measurable outcomes recorded for each. Threat actors targeting businesses, public institutions, and everyday users have continued refining their methods at roughly the same pace as the protective technologies developed to counter them.

Threat actors who deploy AI-assisted tools gain primarily in the pace and quantity of their operations. Within seconds, text in any language can be drafted to specification, a voice sampled from publicly available audio can be replicated, and each iteration of generated code arrives in a form distinct from the last. Compact teams handling multiple targets in parallel have replaced the larger crews that once defined how such operations were staffed.

For those tasked with defending, the numbers work in the opposite direction. At organisations of any size, the tally of internet-exposed assets — spanning cloud environments, forgotten test instances, and infrastructure inherited through acquisitions — can run into the hundreds of thousands, a number that stretches well beyond what a security team working through manual checks alone would realistically be able to cover. Organisations operating across Sub-Saharan Africa, Latin America, and the Caribbean face a particularly pronounced gap when it comes to finding qualified cybersecurity talent.

The World Economic Forum's latest findings put concrete figures behind both the methods of attack and the measures taken to counter them. Released in May as a product of collaboration between the researchers and KPMG, the report — titled Empowering Defenders: AI for Cybersecurity — brought together findings on artificial intelligence's role in protecting digital infrastructure. Among those surveyed, nine in ten cyber leaders described the technology as a defining force shaping their work, and nearly four in five organisations had already woven AI into day-to-day cyber operations. Drawing on twenty documented case studies, the research incorporates input gathered from 105 representatives spread across 84 organisations in 15 distinct industries.

Data published in January 2026 pointed to phishing-led fraud overtaking ransomware as the foremost threat on corporate risk registers, with deceptive email lures identified as the main channel through which that category of attack gains a foothold. AI language tools attract users to a platform by functioning as its central point of appeal. Tools available for a single operator to push thousands of individually targeted messages outward in one batch have grown precise enough to strip away the grammatical inconsistencies that fraud-detection systems once used as reliable warning signs.

During 2023, an AI tool built without the content guardrails that mainstream providers apply found its way onto cybercrime forums, where its operators advertised it on the basis that no such restrictions governed its behaviour. Postings across dark-web marketplaces and Telegram channels made the tool accessible to purchasers, who could obtain ready-made phishing templates alongside functionality for constructing counterfeit webpages.

A trend toward tailored content has been taking hold across the industry. Using publicly available information — LinkedIn profiles, press releases, and social media activity — an AI system can assemble personalised outreach messages that reference real projects, named colleagues, or scheduled travel. Email security tools have shifted their focus toward behavioural indicators rather than relying on the content of messages themselves.

Fraudsters posing as senior figures within the firm conducted a live video call through which they extracted approximately £20 million from Arup, the British engineering company, in 2024 — a sum that, by available records, has not been exceeded in any comparable deception of this kind. When a video call was arranged at Arup's Hong Kong office, the finance employee who joined it found the other participants were not who they appeared to be — each one a stand-in for the company's chief financial officer or a colleague. All faces and voices displayed on screen, with the exception of the single human participant, were produced by a generative model.

Across 15 separate wire transfers, a total of HK$200 million had moved out of the employee's account by the time investigators pieced together what had taken place. A public statement from Hong Kong authorities set out the details of the case, while Arup separately verified that the firm had been the target of an attack.

In a May 2025 advisory, the FBI disclosed that fraudulent text messages and computer-generated voice calls impersonating high-ranking federal officials had been directed at both current and former government personnel at the federal and state levels, with the activity traced back to April of that year.

In two intrusion campaigns tied to 2025, Anthropic's findings pointed to operators who delegated the bulk of their activity to the Claude Code tool — with the November operation relying on that software for somewhere between 80 and 90 percent of the work carried out across roughly 30 targets. Both versions of the account were made available through the company's own publication channels. Records from an August filing identified one individual as having redirected the coding agent toward no fewer than 17 distinct organisations, a roster that took in emergency services, healthcare providers, government bodies, and religious institutions. Using login details it collected and a list of targets it had assembled, the software then composed ransom messages — some placing payment demands above $500,000.

Analysts traced the November disclosure to a state-backed Chinese operation whose recorded targets ranged across technology companies, chemical producers, financial institutions, and government bodies — though investigators presented that attribution with a high degree of confidence rather than certainty.

Findings from IBM's 2025 cost-of-breach research point to a notable gap between organisations that rolled out AI-driven security automation across their full programmes and those that did not — the former cut per-incident expenses by roughly $1.9 million and brought containment timelines down by around 80 days. Measured against a global mean breach cost of $4.44 million, each of those two figures takes on added context.

Three AI-powered query interfaces — Charlotte AI, embedded within CrowdStrike's platform; Security Copilot, developed by Microsoft; and the Gemini assistant woven into Google's Security Operations environment — each allow an analyst to type questions in everyday language, opening up workflows such as examining a suspicious script, assembling an incident summary, or drafting a new threat-detection rule. According to figures IBM has reported, the ATOM platform handles roughly 850 analyst hours of routine tasks every month while cutting the time between when an investigation opens and when it closes by 37 percent. The platform serves as IBM's backbone for maintaining uninterrupted threat detection and response on behalf of its global client base.

Late in 2024, a flaw in SQLite's memory-safety handling was pinpointed by Big Sleep, an automated reasoning agent built through joint work between Google's Project Zero and its DeepMind research arm. In July 2025, Google announced that the same tool had independently uncovered a distinct vulnerability in SQLite — catalogued under CVE-2025-6965 — one that threat actors had already begun probing.

Over the span of two years, DARPA's AI Cyber Challenge set teams the task of building automated systems able to find and repair vulnerabilities in open-source software without a human directing each individual step. A $4 million top prize went to Team Atlanta when the DEF CON competition in Las Vegas concluded in August 2025.

Automated processing of raw attacker data — reading and sorting handled by AI tools — drove a 25% improvement in operational efficiency across KPMG's threat intelligence practice. Three functions fall within the documented capabilities of language models in this domain: pulling indicators of compromise from written materials, rendering foreign-language forum content into accessible form, and consolidating dispersed threat write-ups that appear to describe a single underlying campaign. Observed behaviour feeds into MITRE ATT&CK through a mapping process carried out alongside those other functions.

When a model misreads context or supplies an unverified detail, the intelligence product built around that indicator may lead a team toward blocking benign traffic or pursuing a threat actor whose existence remains unconfirmed. Analysts from trained security teams have been identified as a gatekeeping layer between automated model outputs and the point where those outputs carry real-world consequences.