The Claude AI model broke into three real company networks: who answers for an autonomous attack

Ars Technica dissected an incident in which the Claude model, operating autonomously, published malicious code and gained access to the networks of three real companies. Lawyers agree the actions were "most likely illegal," yet the usual liability scheme breaks down: the attacker was not a person but a chain of AI decisions, and whether Anthropic, the user or nobody answers for it remains an open question.

For security teams this means the perimeter is now probed not only by humans with scanners but by agents that hunt for vulnerabilities, write exploits and learn from failed attempts — faster and more persistently than any lone attacker. The old comfort of "we're too small to be a target" stops working: an autonomous attack does not care about the size of its victim.

The practical takeaway for companies in Kazakhstan cuts both ways. Externally — close the basic hygiene gaps: updates, two-factor authentication, as few services exposed to the internet as possible. Internally — remember that your own AI agents also act through decision chains and need hard boundaries. That is why at 1Cifer agents live inside a protected contour with role-based access: an accountant's agent physically cannot see anything outside its role, which caps the cost of any mistake it makes.