AI is now running deeper inside attacks with less human direction. The old tools for spotting the most dangerous actors can’t keep up.
Hackers are using AI to penetrate compromised systems more deeply, and chain together attack stages with minimal human input. That combination of deeper access and more autonomy is making the standard signals security teams rely on to identify dangerous actors unreliable. Anthropic published those conclusions in a new report analyzing 832 accounts it banned for malicious cyber activity between March 2025 and March 2026.
AI is lowering the bar for getting in
Until recently, writing malware required real programming skill. AI has changed that. In Anthropic’s dataset, 67% of the 832 banned actors used AI to write malware. That matters because attackers who previously lacked the technical ability to build intrusion tools can now get them from a chatbot.
Once inside, AI is taking them further. AI use for account discovery rose 8.9% over the study period. Account discovery is how hackers map out which accounts exist inside a compromised network to find the credentials and access points they need to move deeper. At the same time, AI-assisted phishing, used to gain initial access, fell 8.6%. Attackers are spending less AI effort getting in and more AI effort going deeper once they’re there.
The result: the share of attackers classified as medium risk or higher jumped from 33% in the first six months of the study to 56% in the second.
Why it’s now harder to tell the dangerous ones from the rest
Security teams traditionally assess risk by counting how many distinct attack techniques an actor uses and what tools they access. Both signals have broken down.
In the dataset, the least-skilled actors averaged about 16 distinct techniques. The most skilled averaged about 20. Four techniques separate the bottom from the top. The platform used (Claude Code, API, or standard chat) also showed no correlation with an actor’s perceived danger.
What separates the highest-risk actors is whether they have built systems that let AI run the attack automatically, chaining together stages such as account discovery, lateral movement, and privilege escalation with minimal human direction at each step. Less-skilled actors are increasingly doing this too, which explains why the proportion of medium-risk-or-higher attackers nearly doubled in a single year.
That type of automated, multi-stage attack has no classification in the MITRE ATT&CK framework, the standard reference security teams use to track how attackers operate. Anthropic says it is in discussions with MITRE about updating the framework and has added safeguards to its most capable models to detect and block activities such as malware development and mass data exfiltration.
