AI-powered cyber threats

State of the art (2026)

The threshold moment was Anthropic's November 2025 disclosure of GTG-1002, a Chinese state-sponsored group that turned Claude Code into an autonomous espionage operator against roughly thirty targets — the model executed an estimated 80–90% of the campaign with human operators intervening for only minutes. That collapses the old comfort that AI merely advises attackers; it now runs the operation. The commodity layer has moved in parallel: agentic tooling automates reconnaissance, payload delivery and privilege escalation, the bulk of phishing is now AI-generated, and deepfake voice and video impersonation has become routine in fraud after the Arup case. Defensive AI — XBOW-style autonomous pen-testing, Microsoft Security Copilot, agentic SOC tooling — is real but still trails offensive capability, leaving enterprises exposed during the lag.

Core thesis

Attackers now use AI to do much of the work of a break-in: finding weaknesses, writing malicious code and moving through a network. Anthropic says a Chinese state group used its Claude Code tool to carry out 80 to 90% of a 2025 spying campaign against about 30 targets, with people stepping in at a handful of decisions in each attack. People still choose the targets and decide what to do with what is stolen. An attack that runs from start to finish with nobody directing it has not been shown in public. AI helps attackers sooner than defenders because finding a flaw is faster than getting every affected organisation to fix it.