Latest / Decrypted: The UK Cyber Briefing / An AI agent wiped Azure storage in seven minutes, using a password left on GitHub
Transcript
- 0:03This is Decrypted. Monday, the twenty-eighth of September.
- 0:07An AI agent wiped Azure storage in seven minutes, using a password left on GitHub.
- 0:13Microsoft has detailed an attack where an AI-driven ransomware operator emptied part of an Azure tenant.
- 0:20And the way in wasn't clever.
- 0:22In a write-up published on Friday, Microsoft says an actor it tracks as Storm thirty-one sixty-eight, also ca
- 0:30lled JADEPUFFER, used two stolen software identities.
- 0:34It mapped a victim's cloud for about fifteen hours.
- 0:38Then it deleted storage accounts in roughly seven minutes.
- 0:42Microsoft did not confirm any data theft or ransom demand.
- 0:46The lesson is older than the AI.
- 0:48A secret left in public, and an identity trusted with far too much.
- 0:53Next.
- 0:58The identities were Azure service principals.
- 1:02Think of them as staff badges for software.
- 1:05No phone to approve a prompt, no shift pattern, no lunch break.
- 1:09Microsoft says one spent about 15 hours and 30 minutes on more than 300 successful read operations.
- 1:17It was listing virtual machines, subscriptions and resource groups.
- 1:22A second badge from the same tenant then did the damage.
- 1:25Over 100 storage account deletion attempts in about seven minutes, plus moves against the protections guardin
- 1:32g backup and recovery. Thirty minutes later came more than 30 successful requests for storage account keys.
- 1:39The badge itself, on Microsoft's account, was left on a public noticeboard.
- 1:44An employee had posted the client secret in a public GitHub issue and later edited it out.
- 1:50But the edit history kept it.
- 1:52Microsoft couldn't prove that was the way in, though it calls it the likely vector.
- 1:57And also.
- 2:00Security firm Sysdig documented JADEPUFFER in July.
- 2:05It called it the first large language model driven ransomware operation.
- 2:10Sysdig says the agent got in through an unauthenticated code execution flaw in Langflow, C V E twenty twenty-
- 2:18five, three two four eight, harvested keys, and encrypted 1,342 configuration items in a Nacos service.
- 2:27The encryption key was never stored or sent anywhere, so paying would not have brought the data back.
- 2:33Sysdig also reports a failed login diagnosed and fixed in 31 seconds, and payloads that narrated their own re
- 2:42asoning.
- 2:46Five tokens were issued for the service principal, and two of them were used for deletion within the same 70-
- 2:52second window. Whether a model or a good script sat behind it, the defender's view is the same.
- 2:58Hours of quiet reconnaissance, then a burst too fast for a human to answer.
- 3:03Your only real window is the reconnaissance, when nothing has yet broken.
- 3:08And finally.
- 3:12Three design decisions would have blunted this.
- 3:15First, secrets never live in places people can edit.
- 3:19Prefer short-lived credentials and managed identities, scan repositories, and treat any exposed secret as bur
- 3:27ned, because deleting the text doesn't delete the history.
- 3:30Second, no single identity should both read the estate and delete it.
- 3:35One badge that can list, delete and reach the backup locks is the flaw, whoever holds it.
- 3:41Third, guardrails must not depend on the identity being honest.
- 3:46Microsoft says resource locks and storage account deletion protection stopped deletion of some accounts, even
- 3:53against a compromised identity with broad administrative permissions.
- 3:57The cost is friction. A second approver, a slower legitimate teardown.
- 4:03That is the point. For UK boards, the question is not whether AI attackers are coming.
- 4:09It is whether the account that can delete production can also delete the backups.
- 4:13Opinion, not advice.
- 4:17The written version and every source are linked in the show notes.
- 4:21Thanks for listening.