Seven Minutes and Twenty-Four Hours

Two numbers arrived in the same week, and together they redraw the response timeline for anyone running cloud infrastructure. Seven minutes is how long the destructive stage took in one Azure tenant hit by the actor Microsoft tracks as Storm-3168, also called JadePuffer. Twenty-four hours is how long a China-based actor needed to begin exploiting a freshly disclosed authentication bypass in Rejetto HFS. One number describes how fast an AI-driven attacker can finish a job. The other describes how fast the same class of attacker can start one.

Microsoft describes the JadePuffer operation as the first documented case of agentic ransomware, with a language model driving reconnaissance, credential theft, and destruction. In the tenant Microsoft examined, the destructive stage wiped more than 100 Storage Accounts, plus Key Vaults, SQL databases, Function Apps, virtual machines, and Site Recovery locks. Initial access came from service-principal credentials that had been leaked in a public GitHub issue. Nobody phished anyone and no zero-day was involved. A secret sat in a place indexed by the entire internet, and a model with tool access did what a patient human operator would have taken days to do.

The Rejetto case is the other end of the clock. The security firm Horizon3 used Anthropic's Mythos to find CVE-2026-61500, an authentication bypass caused by Math.random() seeding the keys that sign session cookies. A China-based actor was exploiting vulnerable hosts within a day of disclosure, and The Register reports it as the second Anthropic-linked vulnerability seen exploited in the wild. The fix is HFS 3.2.1 or later. The detail I keep thinking about is the root cause. A weak random number generator in a signing path is a flaw a careful human reviewer finds on a good day. A model found it, and the window between a finding and a working exploit collapsed from weeks to hours.

Two more items from the same fortnight belong beside those. GitLab disclosed CVE-2026-90970, rated 9.9, a prompt-sandbox escape in its self-hosted AI Gateway affecting Duo Agent Platform users on versions 18.1.6 through 19.4. South Korea's president ordered a probe after seven financial firms were breached, with traces of an open-source Chinese-language autonomous pentest framework called ARTEX AI. The first shows that AI plumbing is ordinary server software with ordinary patch obligations. The second shows that the offensive tooling is public and free.

Put those four facts together and the staffing model for incident response starts to look mismatched. An on-call rotation assumes that a page arrives, an engineer wakes, an investigation runs, and a decision follows, a loop that takes tens of minutes at best. When the destructive stage fits inside seven minutes, the loop finishes after the damage is done. Detective controls still matter for the post-incident record and for scoping. They cannot be the primary line when the attacker's cycle time is shorter than the defender's.

That shifts weight toward controls that act before the first request. For the Azure pattern in particular, the details that stand out are specific. The attacker removed Site Recovery locks, which means the principal it held could delete the very protection meant to stop deletion. Long-lived service-principal secrets are the original sin in the story, and workload identity federation removes the secret that leaked. Push protection on public repositories closes the exact door the credentials walked through. Privileged Identity Management on destructive roles means a compromised principal holds eligibility with the standing access removed, and the activation step adds a human gate that an automated attacker has to defeat or wait out.

The patch side follows the same arithmetic. A vulnerability management policy that sets a thirty-day window for high-severity findings was written for an adversary who needed weeks to build an exploit. A window measured in days for internet-facing systems, with a pre-agreed emergency path for anything touching authentication, matches the observed twenty-four hours more honestly. Self-hosted AI infrastructure belongs on that list. An AI gateway sits at the junction of model access, internal data, and credentials, and I read a 9.9 in that position as carrying the same urgency as a 9.9 in a VPN appliance.

Both numbers trace back to the oldest items in the catalog: leaked secrets, removable locks, weak randomness, and unpatched servers. What changed is the speed at which they get chained, which converts every small hygiene gap from a long-tail risk into a near-term one. The attacker's cost per attempt has dropped toward zero, so the gaps get tested continuously, and the only configuration that survives continuous testing is one with no standing gaps.


Sources: BleepingComputer, "JadePuffer agentic AI attacks target Azure" (Oct 1, 2026); The Register on Mythos and the Rejetto HFS flaw (Oct 3, 2026); The Hacker News on GitLab AI Gateway CVE-2026-90970 (Oct 2026); Korea JoongAng Daily on AI-driven bank hacks (Oct 2026).