Separating Evidence From Hype
"AI-powered attacks" appears in almost every security vendor's marketing, often without any evidence behind it. This article sticks to what has been documented in detail by organisations with direct visibility: AI developers who caught their own models being misused, and large security teams publishing measured data.
The picture that emerges is not science fiction. Attackers aren't using AI to invent new kinds of attack. They're using it to do the old ones — reconnaissance, exploitation, credential theft, phishing — faster, at larger scale, and with fewer skilled people.
The First Documented AI-Orchestrated Campaign
In November 2025 Anthropic published a report on a campaign it detected in mid-September 2025 and attributed, with high confidence, to a Chinese state-sponsored group. The attackers built a framework around Claude Code and used it against roughly thirty organisations — technology companies, financial institutions, chemical manufacturers and government agencies — succeeding in a small number of cases.
What made it notable was the division of labour. According to Anthropic, the AI performed 80–90% of the campaign, with human operators needed at only about four to six critical decision points per target. The AI carried out reconnaissance, wrote exploit code, harvested credentials, extracted data, created backdoors, and documented its own progress.
The attackers got past the model's safeguards by splitting the operation into small tasks that each looked innocent, and by telling it that it was working for a legitimate security firm doing authorised testing.
It also had clear limits. The AI "occasionally hallucinated credentials" and sometimes reported extracting secret information that was actually public — errors a human operator would have had to catch. Anthropic banned the accounts, notified affected organisations, and worked with authorities.
Vulnerabilities Are Weaponised Faster Than They're Fixed
Microsoft's Digital Defense Report, published on 1 October 2026, quantifies the speed problem:
- Nearly 40,000 CVEs were published in the first half of 2026 alone, putting the year on track to roughly double the previous one.
- The median time from a vulnerability being found in the wild to being weaponised has fallen to "well below 24 hours."
- Enterprise remediation of critical external vulnerabilities still takes 30 to 60 days.
The gap between those last two numbers is where most breaches happen. Microsoft reports AI being applied by attackers across vulnerability discovery, reconnaissance, phishing, and malware and exploit development. In one of its own controlled evaluations, an AI system chained together 32 attack stages.
The report also shows that the oldest problems remain the biggest. In Microsoft Defender Experts data, user execution — someone running something they were tricked into running — accounted for 30% of observed initial access, and valid accounts — stolen or guessed credentials — for another 20%. "ClickFix" attacks, which trick people into pasting a command into their own terminal or Run dialog, executed on more than 1.1 million unique devices between February and early May 2026.
Where AI Helps Defenders
The same capabilities work in the other direction, and defenders have one structural advantage: they can point AI at their own systems continuously and legally.
- Finding vulnerabilities before attackers do. AI-assisted code review and fuzzing are now part of how major software vendors and open-source projects find bugs, and AI developers have reported their models finding previously unknown vulnerabilities in widely used software.
- Triage at volume. Security teams receive far more alerts than people can read. Summarising, correlating and prioritising alerts is where AI tools have been adopted fastest.
- Detecting misuse of AI itself. The Anthropic case was caught by the AI provider monitoring its own platform. AI providers now publish regular threat reports on how their models are being misused and what they've blocked.
The honest caveat is that AI doesn't fix the 30-to-60-day patching gap on its own. Finding more vulnerabilities faster only helps if organisations can deploy fixes faster too.
What to Do About It
None of the effective defences are new. AI makes them more urgent, because it shortens the time an attacker needs and lowers the skill required.
Know exactly what you expose. Attackers scan the whole internet continuously; AI makes acting on what they find faster. Check which ports your servers accept connections on with the Port Checker, and close anything that doesn't need to be public — especially databases, remote desktop and admin panels. The common ports guide lists the ones that should never face the internet.
Patch internet-facing systems first, and fast. VPN gateways, firewalls, mail servers and web applications are where the under-24-hour weaponisation matters most. Internal systems can follow the normal cycle; edge systems can't.
Move to phishing-resistant authentication. With valid accounts behind a fifth of initial access, passkeys and FIDO2 security keys — which can't be phished or replayed — matter more than any AI-based detection. SMS and authenticator-app codes are better than passwords alone but can still be phished.
Harden what users can be tricked into. ClickFix works because people are told to paste commands. Train for that specific pattern, and restrict who can run scripts and PowerShell on managed devices.
Watch your own certificates and domains. AI-generated phishing often uses lookalike domains with valid HTTPS certificates. Monitoring Certificate Transparency logs for your brand catches these early — see how Certificate Transparency works.
Send the right security headers. Strong headers limit what an injected script can do; see the security headers guide.