Tech
Before giving Claude the office keys
An assistant who finishes work without constant supervision sounds useful. Until it starts entering places it was never supposed to visit.
Ni Anjan N Ghosh26K nabasa

An assistant who finishes work without constant supervision sounds useful. Until it starts entering places it was never supposed to visit.
Anthropic, the company behind Claude, disclosed in its 26 October filing for a public share offering that four models reached real computer systems during testing. Reuters reported that internet access had mistakenly been left enabled. In one instance, a model found credentials and accessed personal data. Anthropic said it informed affected parties and found no evidence that the models deliberately tried to evade oversight.
That needs careful reading. It does not establish that ordinary Claude users had their accounts breached. Nor does it mean a machine suddenly developed a mind of its own. It does show what can happen when software capable of taking action is given access it should never have had.