Nearly two-thirds of the accounts Anthropic has banned over the past year were being used to prepare cyberattacks, the AI company said in a new analysis of AI-enabled threats. The finding comes from a review of the company’s own security logs and covers a 12-month period ending this summer.
What the analysis found
Anthropic said that about 67% of accounts it blocked were linked to activity that aimed to use its AI models to plan or execute attacks. The company didn’t name specific threat actors but described the users as "malicious actors" who are adopting AI at a fast clip. The analysis didn’t break down the banned accounts by region or group, but it noted that the attacks ranged from reconnaissance to writing phishing emails faster.
A jump in attack severity
The broader trend, according to the report, is that attackers are getting more effective. Anthropic’s team found that the severity of attacks is rising quickly as malicious users figure out how to weaponize AI. The company said this isn’t a tomorrow problem — it’s already happening. While the numbers are still small compared to overall usage, the shift in capability matters because even a handful of sophisticated attacks can cause serious damage.
How Anthropic is responding
The company said it’s using the data to improve its safety systems. That includes refining its monitoring tools and updating the rules that trigger an account ban. Anthropic also called for industry-wide sharing of threat data, though it didn’t announce any new partnerships. The findings are part of a broader push by AI firms to get ahead of misuse without stifling legitimate use.
The open question
One thing the report doesn’t answer is how many attacks succeed despite the bans. Anthropic didn’t say whether any of the blocked accounts had already caused harm before being shut down. The company said it will continue tracking the trends and plans to release updates periodically.
