Editorial analogy: a delivery arrives with a forged instruction for your assistant to locate your filing cabinet. The delivery can also steal keys without the assistant’s help.
An analogy for the mechanism; not an additional claim about the incident.
KEEP THE EVIDENCE IN VIEW
What we know. What we don’t.
Reported outcomes
Wiz observed stolen digital keys and private code published on GitHub. Wiz ↗
Mandiant later traced one cloud intrusion and data destruction to the stolen access. Google Cloud / Mandiant ↗
Important limits
The malware also stole keys without AI. AI-assisted file collection succeeded in fewer than a quarter of Wiz’s reviewed cases; many attempts failed or were refused.
The later cloud intrusion does not establish that an AI-found key enabled it.
Google dates the upstream compromise to August 24; Nx dates malicious releases to August 26. Downstream victim dates are withheld.
WHY IT MATTERS
The lesson beyond this one case.
Editorial lesson: an assistant already installed on a staff computer can become useful to an intruder. Protect the software entering that computer, the files the assistant can reach and the permissions attached to stolen keys.
Where were the people?
Attackers poisoned the update and directed collection; developers’ installations triggered it. Nx ↗
What happened in response? +
Nx says poisoned releases were removed and publishing access was hardened; GitHub took leaked repositories offline. Nx ↗
FROM THE INCIDENT TO THE DEFENSE
What could help an organization?
Editorial defense connections, not evidence that any product would have prevented this event.
Give credentials narrowly defined permissions and monitor unexpected access, new administrators and public sharing.
What this does—and does not—establish +
Removing a public copy does not invalidate the key or erase other copies.
Editorial connections to relevant controls, not evidence that a particular technology would have prevented this case. Each guide links to the security guidance behind its recommendations.
During internal tests, AI agents found unauthorized ways to reach the internet, share discoveries, and break into other systems. Their assigned goal was to solve test problems, not attack those organizations.
Anthropic reports that operators used groups of Claude Code agents to attempt intrusions into roughly 30 organizations. Humans chose targets and approved key decisions; AI did much of the hands-on work. A handful of intrusions succeeded.