Anthropic Finds 67% of Banned Accounts Used AI in Attacks

More than two-thirds of accounts banned by Anthropic for policy violations over the last year used AI to help them prepare for cyberattacks, such as writing malware, according to the AI firm. 

Anthropic said on Wednesday that between March 2025 and March 2026, out of 832 accounts that it examined for violating its policies, 560 accounts were used in this way. 

The data reflects an alarming global trend — that AI is increasingly being used to carry out mass cyberattacks. In April, the value of crypto stolen in hacks surged to $629.7 million, the highest since February 2025, which some analysts linked to the widespread use of AI. 

Source: Anthropic

Manuel Aráoz, the founder of the crypto security platform OpenZeppelin, said on May 27 that he considered “all of DeFi unsafe” due to AI models’ ability to identify smart contract vulnerabilities.

While the data shows that most of the AI use is in the preparation phase of an attack, Anthropic said it has also started to be deployed “deeper in the attack life cycle,” with 6.5% of the banned accounts using AI to assist with “lateral movement” — referring to techniques a cyberattacker uses after gaining initial access. 

“These sorts of ‘post-compromise’ techniques used to be restricted to actors with the technical knowledge to carry them out,” Anthropic said. “Our investigation shows that AI can now be made to perform these activities on behalf of less sophisticated actors.”

AI also increased the threat level of attackers. Anthropic classified a third of accounts, or 33%, as “medium risk or higher” in the first six months of its analysis, but that figure nearly doubled to 56% in the second six-month period of its study.

The type of threat posed by AI-powered hackers was detailed by Google researchers last month. The researchers found what they believed was the first-ever case of AI being used to develop a zero-day exploit, which allowed hackers to bypass the two-factor authentication of an unnamed “popular open-source, web-based system administration tool.”

Related: AI guardrail removals raise questions over limits of open-source model regulation

It added that AI can now undertake highly technical tasks for attackers, and there is “little correlation between the skill of a threat actor and how many techniques they use,” a metric that traditionally measured an attacker’s risk level.

Anthropic said in some cases, such as one in November, a Chinese state-sponsored group carried out an attack where an AI model worked autonomously, where it conducted an exploit, stole credentials and made decisions with a human making an input at “key moments.”

“These are precisely the behaviors we expect to see much more of as AI agents become more capable,” it said.

Anthropic is set to roll out its AI model Mythos in the coming weeks, the company’s large language model that has concerned analysts due to its powerful cybersecurity capabilities that found over 10,000 major vulnerabilities in widely-used software.

Magazine: AI-driven hacks could kill DeFi — unless projects act now