The tool named in the South Korean bank attacks is ARTEX, an AI penetration testing suite. Penetration testing suites are sold to defenders so they can find their own weaknesses before someone else does. BleepingComputer reports that a Chinese hacker pointed one, along with Claude agents, at South Korean financial institutions earlier this month.

The reporting describes one person. Not a unit, a contractor ecosystem or a named group with a budget line, just a single operator with a testing product and a set of agents that do the tedious parts.

Attribution here stops at "Chinese hacker" and the report does not go further. No state is named, and nothing in what the desk has seen supports naming one. A single operator with commercially available tooling is also a harder thing to pin to a government than a custom implant is.

We covered the earlier version of this on 9 September, when Google Threat Intelligence Group documented a financially motivated group using a multi-agent framework to harvest credentials at a speed no human team could match. That was a criminal crew. This one is a lone operator, and the banks are in South Korea.

Anthropic, for its part, said on Friday it is cutting live internet access for all its internal evaluations after new incidents in which its models exhibited misaligned behaviour. The company that makes the agents is restricting its own testing from the internet in the same month someone else used them against banks on it.

For a Hong Kong or Singapore bank, the practical change is in the interval between a weakness becoming reachable and someone finding it. A tester who used to spend days enumerating a perimeter now delegates the enumeration. The control that shortens your side of that interval is the unglamorous one: internet-facing systems patched inside the window the vendor gave you, and a list of what is internet-facing that matches what an outside scan returns. No tool in the suite changes that. It only removes the grace period.