Between November 2025 and November 2026, AI labs and security vendors published a steady run of cases in which attackers used AI models in real operations. The reports agree on one point: AI has mostly made known methods faster and cheaper, not created new ones. ENISA said the same in September 2026 and expects more attack stages to be handled by AI. As of November 13, 2026, no public report shows AI was used in the European refrigeration attacks, and UK officials said advanced AI was probably not needed for them.
Documented Cases
| Reported | Reported by | What was claimed |
|---|---|---|
| Nov 13, 2025 | Anthropic | A group it assesses with high confidence to be Chinese state-sponsored used its Claude model in a largely automated campaign against roughly 30 organisations, succeeding in a small number. Anthropic estimated AI carried out 80 to 90% of the work |
| Feb 3, 2026 | Sysdig | In a cloud intrusion on November 28, 2025, an attacker went from first access to administrator rights in under ten minutes. Sysdig inferred AI assistance from the code's style and errors |
| Mar 6, 2026 | Microsoft | North Korean groups used AI to build fake identities for remote IT jobs, alter documents and write malware. Microsoft said it had disrupted thousands of related accounts |
| May 11, 2026 | Google Threat Intelligence Group | Criminals planned mass exploitation using a previously unknown flaw that Google believes, with high confidence, an AI model helped find. Google alerted the vendor and disrupted the plan |
| May 11, 2026 | Google Threat Intelligence Group | State-linked groups used AI for vulnerability research, and four malware families used AI-generated code to hide their function |
| Sep 2026 | Anthropic | A group it assesses as Russian state-linked automated nearly every stage of espionage against more than 20 organisations, concentrated in Ukraine and Europe |
| Sep 2026 | Anthropic | One French-speaking hacktivist targeted 42 European political parties, media outlets, think tanks and their software providers, gaining internal access in at least 14 |
| Sep 2026 | Anthropic | A Chinese-speaking espionage group targeted roughly 50 organisations and ran automated vulnerability research that found previously unknown flaws in a security product |
| Nov 12, 2026 | Symantec | Attackers on a US retailer's network in September 2026 switched to AI-written scripts after their first attempts were blocked. Symantec says it has investigated dozens of AI-assisted attacks since May |
Each row is the reporting company's own assessment. AI labs can only see activity on their own platforms, and vendors infer AI use from traces such as code style. The attributions to states are theirs and have not been confirmed by governments in these reports.
What the Reports Agree On
- Acceleration, not invention. Microsoft calls AI an accelerator with humans still making decisions. Symantec says AI supplied no new techniques. Anthropic says most techniques were familiar and the change is in speed, scale and cost.
- Agentic use is early. Microsoft said in March 2026 that it had not seen agentic AI used at scale. The September Anthropic cases suggest that changed during the year for some actors.
- Models make mistakes. Anthropic noted invented credentials. Sysdig and Google both used AI errors as evidence that AI was involved.
- Europe is a target. Two of Anthropic's September cases centre on European governments, parties and media.
ENISA's Assessment
ENISA addressed AI in two 2026 publications. Its Threat Landscape 2026 finds that in 2025 attackers mainly used consumer AI tools to improve existing skills. It adds that advanced models released in 2026 have already shaped the threat landscape, and it assesses that AI will highly likely support more malicious operations, with 2026 likely to see more attack phases directly enabled by AI. See ENISA Threat Landscape 2026 statistics.
Its paper of July 7, 2026, on cybersecurity in the frontier AI era, is about speed.
| Figure cited by ENISA | Value | Origin |
|---|---|---|
| Time to weaponise a newly disclosed vulnerability | Within 15 minutes | Industry research cited by ENISA |
| Median time from first access to data theft | 72 minutes | Industry research cited by ENISA |
| One organisation's vulnerability identifiers per quarter | About 80 in early 2025, close to 500 in early 2026 | Example given by ENISA |
| Same organisation with frontier AI tools | About 500 reports a day | Example given by ENISA |
| Breaches where logs existed that should have raised the alarm | 75% | Research cited by ENISA |
These are figures ENISA quotes from others, not its own measurements. The paper describes the gap between a flaw being found and being exploited as moving from years to months to, potentially, hours or minutes. It recommends putting effort into prioritising and fixing flaws over finding more of them, and automating defence.
Where AI Was Not Needed
The refrigeration attacks on European supermarkets in October and November 2026 caused more physical disruption than any case in the table. Officials at the UK National Cyber Security Centre said advanced AI was probably not required, because the attacks used known weaknesses in old equipment. The centre's head, Richard Horne, said AI will make such weaknesses easier to find and exploit at scale. The centre's annual report rates an AI-powered ransomware attack on UK critical infrastructure by 2028 as highly likely. The incidents are covered in our refrigeration attack timeline and the wider year in the Europe cyberattack tracker.
Methodology
Cases were taken from the publishers' own reports: Anthropic, November 2025; Anthropic, September 2026, of which we read the first part; Sysdig; Microsoft; Google Threat Intelligence Group; and Symantec. ENISA material is from the Threat Landscape 2026 and ENISA's view on cybersecurity in the frontier AI era. The UK remarks are from an Associated Press report carried by Insurance Journal. We could not open OpenAI's 2026 reports, so they are not in the table. Incidents are described at news level only. Read November 13, 2026.