Anthropic: AI boosts low-skilled hackers’ cyberattack capabilities

Anthropic’s Frontier Red team analyzed 832 accounts banned between March 2025 and March 2026 and found that AI enabled low-skilled actors to plan, build or execute parts of cyberattacks that previously required higher technical skill. The review covered cases where account holders used Anthropic models and tools to assist attacks.

The study shows the share of actors rated medium risk or higher rose from 33% in the first half of the year-long period to 56% in the second half, a 1.7-fold increase. Among the accounts, 67.3% used AI to assist in malware development and 6.5% used it for lateral movement inside compromised systems. The least-skilled actors averaged about 16 distinct techniques while the most-skilled averaged about 20. The platform type-coding models, API integrations or chat tools-did not reliably predict how risky an actor would be.

Anthropic measured AI’s role across the attack life cycle. Use of AI in phishing fell by 8.6%, while AI-assisted account discovery within compromised networks increased by 8.9%. The report documents AI supporting privilege escalation, lateral movement and other operationally demanding steps.

The Frontier Red team wrote that “Now that AI can perform highly technical tasks on an actor’s behalf, there’s little correlation between the skill of a threat actor and how many techniques they use.” The team added that where actors apply AI in the attack chain has been a distinguishing signal, but that signal is weakening.

The report highlights risks for the cryptocurrency industry, where exchanges, protocols and digital wallets face complex, high-value attacks. The sector recorded 40 major hacks in May 2026.

Anthropic’s report does not attribute the increase to any single model or provider and emphasizes behavioral signals over tool-based attribution. The analysis covers activity the company observed on its own systems and may not reflect the full extent of AI-assisted cybercrime on the wider internet.

The report notes that security teams have traditionally judged threat levels by counting techniques or identifying tools, and that metric can be less reliable when AI automates complex steps.