Every SOC has the same worry right now after ARTEX. “Did AI hyperattacks just become an app for attackers?” We have been looking at hyperscale attacks and the tooling behind them for a while. Last Wednesday we published what we learned from putting a live production environment in front of 26,000 AI agents for three days: 231 billion events, 101,169 alerts, and one event in 13,338 that was an attack. The attack was not hidden. It was buried.
Recently, a trusted cloud application, carrying a credential stolen from its vendor, attempted to sign in to more than 2,250 Microsoft 365 tenants. Huntress reported that in some cases, the attackers were able to access global administrator credentials within minutes of initial access. The techniques were common to security practitioners, but the scale and the speed were not.
On November 13, 2026, the FBI issued a FLASH warning that ARTEX attacks are ongoing. This confirms what many of us suspected and what TENEX.AI demonstrated in our Hyper-Attack exercise: autonomous attack frameworks are increasingly accessible, already in use, and moving faster than traditional defenses can respond.
This post covers what changed, why the detection fundamentals still hold, and what is worth doing this week.
ARTEX is an open-source AI penetration testing framework that was released on GitHub in July 2026. It has a planning agent that maps out the target and assigns work to various worker agents. The worker agents then run the standard offensive tools and report back until the objective is achieved. It was published on GitHub on 26th July 2026 by a developer known as Autumn-27 and is compatible with almost any model. The operators seen so far have all favored the cheaper models. It had over 12,000 forks when the author went closed source and there are still archived copies, translations and rebranded versions of it out there. Censys reported 1,465 exposed ARTEX consoles in early November (about 815 distinct hosts once identically configured blocks are grouped).
It's already been seen in the wild. Microsoft has already attributed an October supply chain campaign across a range of Microsoft 365 tenants to a group it tracks as Storm-0940. CrowdStrike has seen a separate China-nexus actor (VAPOR PANDA) using it in an 11-day window to attack over 200 organizations across 14 countries. In October, attacks on South Korean banks that exposed customer data and caused outages were linked to ARTEX via the attacker's own exposed config and session files. Microsoft, CrowdStrike, Cisco Talos, Zscaler, and the FBI all came to the same conclusion: the same techniques, just at a much faster pace.
How the threat works
The October campaign is the clearest example, and Microsoft and Huntress have published the sequence.
• October 10 to 14: a single address conducted a password spray of 327 accounts across eight organizations. More than 151,000 failures, peaking at 2,900 a minute, and zero successes. It didn't need to work.
• The credential for the multi-tenant backup application was stolen at the vendor. The “how” is not public.
• On October 15, agents signed in as that application across more than 2,250 tenants, moving through them in sorted tenant-ID order a few seconds apart. Huntress saw sign-ins against just under 1,000 tenants from one address, with more than 300 successful. Microsoft reports that in a number of instances the time from initial targeting to privileged access was under three hours.
• In some customer tenants, the app had assigned itself Global Administrator and then within about five minutes had gone on to add client secrets to other apps with expiry dates in 2028.
• The follow-on activity involved directory reconnaissance via Graph, Azure resource enumeration, bulk reading of secrets across the key vaults and access to mailboxes and SharePoint.
• On October 17, the vendor disabled the apps in question and suspended the service. The planted secrets and role grants survived that.
Three things stand out and none of them is a new exploit. The actor is a non-human identity with the permissions it was given. The front door is the supply chain, not a phished user. The clock is in the minutes: admin rights to persistence in about five, with a kill that doesn't undo it.
Why it matters for defenders
This is not a new kind of attack. It is an enhancement of automated attacks we have watched for years, with one addition: the AI agents troubleshoot as they go. The same actor is the cleanest example. Before ARTEX, Microsoft tied Storm-0940 to the Quad7 botnet, thousands of compromised routers running password sprays against Microsoft 365 at scale since at least 2021. That was automation: fast, broad, and dumb. When a spray failed, it kept spraying. In October, the same group ran a spray that failed 151,000 times, then moved on to a stolen vendor credential, figured out what each tenant allowed, and adapted the next step based on what it found. That is the difference. The automation now replans.
It also means everything in the public reporting is disposable. The IP ranges, the user agents, the application IDs. Wiz saw new addresses appear within minutes of others being blocked. ARTEX replans after every step, writes fresh scripts each time, and its operators told it to skip hosts that had endpoint protection. By the time a blocklist ships, the names and addresses have changed.
What does not change is what an intrusion has to do. An application signs in from a new place and gets a role. A service principal adds a credential to another application. A non-human identity reads secrets across several vaults in minutes. A single source pulls down source maps and configuration files, then walks every route it finds. Those patterns show up no matter what the tool is called.
Here is the hard part. Attackers, pentesters and red teams have abused service accounts, application and machine identities for years. What changed is speed and scale: machine pace, thousands of tenants at once. e. And this activity has always been harder to catch than a person. Detecting a non-human identity means knowing what it normally does, which means maintaining a baseline for every application and service principal in the tenant, updated as vendors change their automation. Most SIEMs never receive that telemetry, and most programs never reach the maturity to maintain it, so these identities go largely unmonitored. A backup application reads every mailbox as part of its job; without a baseline, volume alone cannot tell you whether it is the app or an attacker holding its credentials. When we checked our own library against this campaign last month, we found exactly those gaps. The challenge is not knowing the behaviors. It is putting the baselines into practice at the speed and scale attackers now work at.
How to defend
Defenders do not need an exotic new tool as a starting point. They need the controls they already own to be complete, connected and fast. Seven steps, in the order we would take them.
- Harden every internet-facing asset, especially secondary login portals and mobile app backends. Rate-limit and alert on bulk record lookups.
- Know your vendor applications. Inventory every multi-tenant and partner application with a foothold in your tenant, what it can do, and what you would do if its credentials were stolen. The October victims did nothing wrong inside their own tenants.
- Govern and baseline every identity, human and non-human. Cloud applications, service accounts, service principals, IoT and machine identities. Limit who can grant application permissions, alert on new application credentials and admin role changes, and record how each identity normally behaves so departures stand out.
- Close the telemetry blind spots. Get logs from applications, servers, network devices, endpoints, security consoles and identity systems into one place, and keep constant watch on the crown-jewel applications and services. Agentic attackers look for the gaps across the stack and go there first.
- Automate containment within agreed authority. Disable accounts, revoke tokens and isolate hosts in minutes when a spray burst, an unusual run of application sign-ins or a web shell appears. A person defines the authority; the response runs at machine speed inside it.
- Retire stale secrets. Remove long-lived keys and VPN credentials or, at minimum, rotate them, review secrets with far-out expiry dates, and watch public code and breach dumps for leaks.
- Test like the attacker. Have a qualified red team or testing partner run authorized AI-driven testing against your own systems, in an isolated environment with no production credentials.
A hard, layered environment does more than slow an agentic attacker down. It makes the attacker noisy. Every denied attempt, every decoy hit, every blocked connection is signal, and that signal is what machine-speed detection and response is built to act on.
Detect behavior, not tool names
These five come from this fall's ARTEX reporting, most of them from the October campaign. The next one will not look exactly like it: different entry point, different identities, new infrastructure. What carries over is the class of behavior: a non-human identity doing something it has never done, in a sequence only an intrusion needs. Treat these as worked examples, not a checklist.
- An app signs in from outside its normal network, then within 24 hours, that same app changes a role or credential. Join on the app's object ID, not the display name. That is the October campaign, and it generalizes: new place, then new power.
- A service principal assigns itself a privileged role or application role. Initiator equals target. Very low volume; review PIM and vendor automation before alerting, and tune by vendor rather than turning it off.
- An app adds a credential to another app. Key on the initiator being non-human, not the target. That is the five-minute secret planting, and it is how persistence survives a vendor's kill switch.
- Bulk secret reads by a non-human identity across multiple vaults. Microsoft's threshold of 20 or more reads across three or more vaults in 15 minutes is a starting point; set your own from a baseline. Watch key unwraps as well.
- Web recon against internet-facing apps: source map, config and Swagger fetches, then route walking and a one-source burst of 404s. Most SIEMs have virtually no rules on web logs. NDR metadata sees it at layer seven without any app forwarding logs.
Three practical notes. Entra audit records rarely contain the initiator's IP, so audit rules have to key on the initiator identity. Graph directory reads generate no audit record; the only surface is the Microsoft Graph activity log, and if it is not collected, that stage cannot be hunted. And every rule above needs a baseline of what each application normally does; that baseline is the work most programs never finish.
How TENEX.AI approaches this
TENEX.AI recommends a defense-in-depth approach across your infrastructure.
- Revisit SIEM capacity and spend now. The economics just changed on the attacker's side, so the risk math and the spend behind it should change too. Year-end budget season is the time to ask whether cloud, identity and web server logs are actually getting in. Keep the SIEM as your system of record.
- Operationalize at scale. The October actor worked through more than 2,250 tenants in order. Coverage has to include every tenant, every application identity and every alert, not just a sample above a severity threshold.
- Get to the evidence where it lives. Graph activity logs, key vault diagnostics, web logs and NDR metadata rarely make it to the SIEM. Collect them, query them in place, or route them. Do not wait for them to arrive.
- Correlate the agent, the network and the identity as one actor. The attacker is an application identity moving across the network. Detection has to see the directory, the network and the workload at the same time.
- Orchestrate on the attack's clock. New intelligence has to become a hunt, an impact check and a tested rule in hours, not on a weekly content schedule, with a person approving every rule before it deploys.
Acting on attacker noise at machine speed is the point. Doing it without accountable people is a new risk, not a defense. Obvious attacks get stopped at machine speed and escalated for validation, and the judgment calls go to people, with the line between the two set by each customer's risk profile. AI handles the volume. Humans handle the judgment.
FAQ
Does defending against ARTEX require new tools?
• Not as a starting point. Every published recommendation from the investigating vendors is a conventional control done well. What changes is the speed at which those controls have to be connected and acted on. When a team cannot investigate every alert or respond overnight, fully-agentic operations with human oversight add capacity.
What should we check first?
• Applications. Inventory every multi-tenant and partner application with admin roles or application-role grants, and hunt from October 1 for role and credential changes where the initiator is not a human. Then turn on Graph activity logs and key vault diagnostics so the next hunt has somewhere to look.
How do we tell ARTEX-style activity from normal automation?
• Not by the tool's name; defenders will rarely see it. By sequence and by identity. A backup application reading mailboxes is normal. The same application signing in from a new network and then granting itself a role is not. Baseline what each non-human identity does, and alert on the sequence, not the volume.
Will the vendor's kill switch fix it?
• No. Disabling the compromised applications stops new sign-ins. It does not remove the role grants or the secrets planted in other applications. Those have to be found and removed.
Sources
• Microsoft Security Blog, Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise, October 21, 2026: https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/ (2,250+ tenants, 2,900 attempts a minute, 327 accounts, eight organizations, 151,000+ failures, under three hours to privileged access; confirmed via CyberScoop Oct 21 and The Hacker News Oct 22 quoting the post)
• CrowdStrike, VAPOR PANDA uses ARTEX to scale global campaign, via BleepingComputer, November 2, 2026: https://bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs (200+ organizations, 14 countries, operators told agents to skip protected hosts); and Help Net Security, November 9, 2026: https://www.helpnetsecurity.com/?p=390753 (11-day window, July 26 release, Autumn-27, 12,000+ forks, rebranded copies)
• Censys, ARTEX: Agentic Asbestos, November 5, 2026: https://censys.com/blog/artex-agentic-asbestos (1,465 exposed consoles, about 815 distinct hosts); TechCrunch, November 6, 2026 (800+ deployments reachable without a password)
• Huntress, ConnectWise (SkyKick) vendor credentials abused in M365 attacks, October 16, 2026, updated October 27: https://www.huntress.com/blog/vendor-credential-enables-unauthorized-access (Global Administrator self-assignment at 03:44 UTC October 15, Python-urllib/3.12, secrets added to other apps expiring 2028, sorted GUID order, almost 1,000 tenants and 300+ successful sign-ins, vendor disabled apps October 17)
• The Hacker News, China-linked hackers use ARTEX AI, October 22, 2026 (updated October 23 with Wiz): https://thehackernews.com/2026/10/china-linked-hackers-use-artex-ai.html (Zscaler on DeepSeek, GLM and Grok configurations; Wiz on new IPs within minutes)
• BleepingComputer, Hacker used ARTEX AI and Claude agents to target South Korean banks, October 10, 2026: https://www.bleepingcomputer.com/news/security/hacker-used-artex-ai-and-claude-agents-to-target-south-korean-banks/ (Shinhan, KB Kookmin, Hana named; customer data and card information exposed; ARTEX confirmed by CrowdStrike from exposed config and session files). The 67,000-record and 8.2 million figures are not in this source and are not used.
• TENEX.AI Hyperattack exercise with Armadin, July 2026.

