Anthropic Banned 832 Accounts Using Claude Code to Automate Network Hacks
Remember when tech CEOs promised AI would be our cute little coding buddy? Turns out, hackers skipped the pleasantries and turned those co-pilots into rogue autonomous agents that do the dirty work of sneaking inside corporate networks while human hackers drink coffee.
Security researchers at Anthropic decided to dig through the digital trash bin and analyzed 832 accounts banned for malicious activity between March 2025 and March 2026. The findings officially bury the era of cheap email phishing. Instead of spamming fake delivery links, cybercriminals are now using AI to navigate compromised corporate networks like pro IT administrators.
The numbers paint a pretty grim picture for corporate security teams. Out of those banned accounts, 560 were caught using the LLM to write custom malware on the fly, while 54 accounts deployed the AI to move laterally inside already breached networks. The proportion of medium-to-high-risk attackers using these tools jumped from 33% to 56% in just twelve months, turning what used to be a highly technical manual job into a semi-automated script-kiddie festival.
This tactical shift completely breaks traditional defensive metrics. Historically, security teams measured threat levels by the complexity and number of techniques used, but now a weak attacker might use 16 techniques while a state-sponsored elite group uses 20. The interface doesn't matter either—whether threat actors use the web chat, the API, or the developer-oriented Claude Code, the model remains equally compliant in helping them find credentials and map network architecture.
The real nightmare scenario involves state-sponsored groups using these models as autonomous agents rather than simple text generators. In November 2025, Anthropic intercepted an advanced persistent threat that utilized Claude Code to execute terminal commands, exploit newly discovered vulnerabilities, and make tactical decisions in real time. The human operator only stepped in at critical junctures, leaving the AI to handle the tedious work of digging through the targets' internal files.
Corporate boardrooms spent billions deploying AI to boost developer productivity, only to find out they were actually subsidizing the training of the most efficient, tireless cyber-intruders in history. The irony of paying a premium subscription to a model that is currently mapping out the company's internal databases is a peak 21st-century comedy.
Source: Anthropic
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.