The Agentic Post
Breaking
Black October 2026: Every Major AI Security Incident, Explained  ·  ARTEX: The Open-Source AI Hacking Tool Behind October’s Biggest Attacks  ·  Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts  ·  PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch  ·  OpenAI-Backed Thrive Buys ADT for $15.3 Billion to Rebuild It With AI  ·  The AI Data Center Backlash Hits Minnesota and Australia in One Week  ·  
Home/AI Safety
Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts

Chinese Hackers Used AI Agents to Hit 2,250 Microsoft 365 Accounts

AI Safety

Microsoft says China-aligned Storm-0940 used the ARTEX AI agent tool to target 2,250+ Microsoft 365 tenants after breaching ConnectWise Cloud Backup. Huntress confirmed hundreds of successful logins.

Microsoft says a China-aligned hacking group used AI agents to attack more than 2,250 Microsoft 365 business accounts after breaking into a backup service that IT providers use to protect their clients’ data. The group, which Microsoft tracks as Storm-0940, got in through ConnectWise Cloud Backup, formerly known as SkyKick. Microsoft published its findings on October 21, 2026, and said the AI agents scanned, guessed passwords and abused permissions with limited human involvement.

How did the attackers get in?

Through a trusted middleman. ConnectWise Cloud Backup lets IT service providers back up their customers’ Microsoft 365 email and files. To do that, it holds powerful "application identities", think of them as master keys that can open many customers’ accounts at once. Storm-0940 compromised several of these keys. In some cases it added new authentication keys and changed administrator roles so it could keep its access even if a password was reset.

That is what makes supply-chain attacks so dangerous: the victims never did anything wrong. Their IT provider used a reputable backup tool, and the backup tool became the way in.

How big was the attack?

Large and fast. Microsoft says Storm-0940 tried to sign in to more than 2,250 Microsoft 365 tenants, at up to 2,900 attempts a minute, and in some cases gained high-level cloud access within three hours. Most attempts failed. Security firm Huntress, which raised the alarm first, confirmed breaches at 162 tenants by October 16 and later found nearly 1,000 tenants hit by sign-in attempts, more than 300 of which showed successful logins.

What role did AI play?

The group used ARTEX, a free, open-source hacking tool built around AI agents that can run on different AI models. In plain terms, instead of a person typing every command, the AI agent worked through the steps itself: mapping networks, trying passwords and testing what each stolen key could unlock. Microsoft did not say which AI model powered it. ARTEX is the same tool linked to attacks on Korean banks and targets in 19 countries, and its developer pulled it from GitHub on October 8 citing misuse, too late to stop copies already in circulation.

Was customer data stolen?

Not confirmed. ConnectWise says there is "no indication" backup data was compromised. It disabled seven application identities on October 17, took Cloud Backup offline, and began a staged restoration on October 21 that it expects to take several days. ConnectWise has not publicly responded to Microsoft naming Storm-0940. Microsoft says the incident did not come from a flaw in its own systems.

What should businesses do?

  • Keep ConnectWise integrations disabled until ConnectWise confirms they are safe to restore, per Microsoft’s advice.
  • Review Microsoft 365 sign-in logs going back to at least October 15, 2026, looking for logins from apps or places you do not recognise.
  • Audit app permissions: check which third-party apps hold broad access to your tenant, and remove any you no longer use.
  • Ask your IT provider directly whether they used ConnectWise Cloud Backup or SkyKick.

It is the same lesson as AgentCorruption: one over-powerful key in a shared service can open everything behind it. See our Black October roundup for the full month.

Read The Next Web’s report on Microsoft’s findings.

Up Next
PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch

PHP2Shell Is Now Being Exploited, and Attacks Began Before the Patch

AI Safety

CISA confirmed both PHP2Shell flaws are being exploited. Amazon saw the first attempt four days before disclosure, and researchers logged 68,000+ attack requests from 73 countries. Update PHP now.

The PHP2Shell flaws are no longer a theoretical risk. The US Cybersecurity and Infrastructure Security Agency (CISA) has added both bugs, CVE-2026-109618 and CVE-2026-109619, to its list of vulnerabilities known to be exploited in real attacks, and gave federal agencies until October 29, 2026 to fix them. Amazon says the first exploitation attempt it saw came four days before the flaws were made public. If your site runs PHP and you have not updated, do it now: the fixed versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35.

What changed since the patch came out?

When PHP shipped emergency fixes on October 21, the question was how fast attackers would turn the bugs into working attacks, given that AI tools made the complex exploits far easier to build. The answer: almost immediately. CISA had said the week before that it had not yet seen exploitation. It has now confirmed it.

Delft University of Technology’s threat intelligence lab logged 68,499 exploitation requests from 652 IP addresses in 73 countries since October 21, with 174 of those addresses sending actual attack code. Cloudflare CEO Matthew Prince said his network saw thousands of attempts in the ten hours before disclosure and millions shortly after.

How did attackers get in before the disclosure?

Nobody knows yet. Amazon Threat Intelligence said its MadPot honeypot, a decoy system AWS uses to study attackers, received a PHP2Shell attempt on October 17, four days before the planned disclosure and three days before details leaked on Hacker News. The request came through a commercial anonymising network. Variations in the payloads suggest more than one group was involved. Whether those early attempts would have worked on real servers is unclear.

Have real systems actually been broken into?

Yes, though the details vary in how well they are confirmed. The strongest evidence comes from Adel Karimi, detection lead at Perplexity, who recorded 35 shell commands launched through Apache on a test server on October 24, matching the first flaw. Security researcher Marcus Hutchins called it the highest quality confirmation of real-world exploitation he had seen.

A more alarming claim comes from security firm Dream, which says a chained attack rooted 25 servers, including 24 firewalls and systems in a US government-affiliated environment, by pairing PHP2Shell with a Linux kernel flaw. Dream pointed to a "Chinese-nexus" operation using AI tools including Claude Code, DeepSeek and Kimi, but had no direct proof those tools wrote the exploit. The Next Web, which reported it, said it had not verified the findings.

Not every early report held up. Sophos corrected a report of successful attacks after its data showed only scanning and failed attempts, and Palo Alto Networks’ Unit 42 withdrew its assessment of an attack on a European government agency. That is normal in the first week of a major flaw, and a reason to read breach headlines carefully.

What should site owners do now?

  • Update PHP to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or later. In WordPress, check under Tools, Site Health, Info, Server.
  • If you cannot update, CISA’s advice is to take the system offline. PHP says no configuration change fully protects you, and the risk is highest with Apache mod_php.
  • Check firewalls and appliances that run PHP in their management panels, not just websites.
  • If you were exposed before patching, look for unexpected processes started by the web server and rotate passwords, database credentials and API keys.

This is the pattern Microsoft warned of when it said flaws now get weaponised in under 24 hours. PHP2Shell beat that: attacks started before the patch existed. See our Black October roundup for every major AI-era security incident this month.

Read The Next Web on CISA’s confirmation and Infosecurity Magazine on Amazon’s early detection.