Anthropic released a detailed analysis examining 832 accounts banned for malicious cyber activity between March 2025 and March 2026, mapping them onto the MITRE ATT&CK database. The findings show that preparing for a cyberattack, specifically writing malware, was the most common AI-enabled activity, utilized by 67.3% of the studied accounts. Over the study period, attackers increasingly applied AI deeper into the attack life cycle for post-compromise techniques rather than just initial access. The proportion of actors classified as medium risk or higher increased from 33% in the first six-month period to 56% in the second. Anthropic found that traditional signals for assessing attacker risk, such as the number of techniques used or the interface chosen, no longer correlate strongly with threat level. Instead, higher-risk actors are distinguished by where they apply AI in the attack life cycle and the scaffolding they build to allow models to execute steps sequentially with minimal human input. The report highlights that behaviors like autonomous agent orchestration are not yet included as attacker techniques in the MITRE ATT&CK framework. In response to these findings, Anthropic has deployed cyber safeguards on its models, engaged in discussions with MITRE regarding framework evolution, and is providing employee-like access to METR to evaluate frontier model safety.
What we learned mapping a year’s worth of AI-enabled cyber threats
Anthropic published a new report examining 832 accounts banned for malicious cyber activity between March 2025 and March 2026. The research maps AI-enabled cyber threats against the MITRE ATT&CK framework and highlights trends in attacker sophistication, post-compromise techniques, and autonomous agent use.
Why it matters
This report reveals that AI lowers the barrier for complex post-compromise cyber attacks, challenging traditional security frameworks that rely on skill level or technique counts to measure risk.