How AI is making cyberattacks faster, cheaper and harder to spot

• Cost of persuasion falls. A fluent phishing message in any language, tailored to a target's job, no longer needs a skilled writer. Voice and video can be cloned too: see deepfake fraud.
• Cost of reconnaissance falls. Finding which companies run a vulnerable product, reading its documentation and drafting a working attack against it is the kind of grind AI tools speed up.
• Speed rises. Once inside, attackers move to other systems faster. CrowdStrike's 2026 Global Threat Report, published 24 February 2026, puts the average time from first access to moving sideways at 29 minutes, with the fastest observed at 27 seconds (CrowdStrike's own data on 2025 activity).
• Fewer fingerprints. The same report says 82% of 2025 detections were malware-free, meaning attackers used stolen logins and legitimate tools rather than code a scanner can flag.

CrowdStrike reports an 89% year-on-year increase in operations by AI-enabled adversaries in 2025. That is the company's measurement of its tracked adversaries (it follows 281), not of every attacker. Be wary of the larger numbers that circulate online (phishing “up 1,265%”, “82% of phishing emails are AI-written”): the ones we found trace to aggregator blogs rather than a primary study, so we do not repeat them. A good rule: if a statistic has no named study, sample and date, it is not evidence.

In November 2025 Anthropic published a report on a campaign it attributed to a Chinese state-sponsored group, tracked as GTG-1002, that used its Claude Code tool to attack about 30 organisations including technology firms, financial institutions, chemical manufacturers and government agencies. Anthropic said the AI performed 80–90% of the tactical work, with humans stepping in at a handful of decision points, and that the attackers succeeded in a small number of cases. The Register reported the same.

It was disputed. Critics noted that the report published no indicators of compromise, so defenders could not check whether they were targets, and that the model sometimes overstated what it had achieved. Treat it as a credible, company-reported early example of attackers wiring an AI agent into the attack chain, not as proof of how widespread it is. The same idea underlies agentic AI in general: software that plans and acts through tools. It is as useful to an attacker as to anyone else.

The same speed that helps attackers helps those who patch, triage alerts and read logs. The practical risk is the gap: attackers can adopt AI immediately; most organisations take months to patch or review a device. That gap is exactly the one the supermarket fridge attacks exposed. Belgium's Centre for Cybersecurity said after them that “all IT systems are potentially vulnerable, and AI can now be used to successfully exploit these weaknesses”. The response is dull and effective: shrink the time between a flaw being known and being closed, and remove what does not need to be on the internet.

Teams building their own AI agents should apply the same logic to their own tools: least-privilege credentials and a human sign-off for irreversible actions, as covered in AI agents and Payaries.

• 2026 CrowdStrike Global Threat Report: AI accelerates adversaries and reshapes the attack surface — CrowdStrike
• Chinese spies told Claude to break into about 30 critical orgs. Some attacks succeeded — The Register (13 Nov 2025)
• Anthropic uncovers first large scale AI-orchestrated cyberattack targeting 30 organizations — The Decoder
• New cyberattack on supermarket fridges sweeps across Europe — CNBC