跳到正文
DEV Community· CD Security Lab·· 13 小時前AI 評分71

網絡安全股為何上漲?ARTEX AI 攻擊工具與 Microsoft 365 3 項檢查

ARTEX AI Hacking Tool: Why Security Stocks Jumped + 3 Microsoft 365 Checks

AI 導讀

ARTEX 被用於針對 Microsoft 365 租戶的攻擊,作者整理網絡安全股上漲背景,並提供 3 項租戶檢查。Microsoft 稱相關登入嘗試涉及超過 2,250 個租戶,部分案例從首次鎖定目標到取得特權存取少於 3 小時。檢查涵蓋 ConnectWise 應用程式 ID、服務主體登入及高權限角色變更;作者提醒查詢需按租戶調整,且未能在含相關應用程式的租戶中測試。

正文

In February 2026, one AI tool knocked cybersecurity stocks down. Investors thought AI would replace security companies. Nine months later, an AI attack tool called ARTEX did the opposite: it reminded everyone why security teams exist, and the same stocks jumped.

The short answer: AI did not replace security. It gave attackers speed. And one of the first big ARTEX campaigns went straight into Microsoft 365 tenants, through apps that companies already trusted.

This article is the written version of my video. It has the timeline, the real numbers (I checked every one), what Microsoft says, and three checks you can run in your own tenant in about ten minutes, with the queries.

This is not investment advice. I am a security person, not a financial adviser. Stock prices are here only to show how the market reacted to the threat.

Video chapters

  • 0:00 AI vs security stocks
  • 0:14 February 20: the scare
  • 0:41 Code is one small job
  • 0:52 ARTEX: attackers got AI too
  • 1:23 Microsoft 365: signed in as the app
  • 1:57 Companies are buying more security
  • 2:37 What Microsoft says
  • 3:11 Check 1: find the 3 apps by ID
  • 3:25 Check 2: service principal sign-ins
  • 3:42 Check 3: who got powerful roles
  • 3:54 If you find something
  • 4:02 The lesson

1. February 20, 2026: the scare

News from February 20, 2026.

Anthropic showed a tool called Claude Code Security. In plain words: it reads your code, finds bugs, and suggests a fix. It was a limited research preview.

Investors got scared. The thinking was simple: if AI can do security, who will pay the security companies?

Here is what happened that day (closing price against the day before):

Stock One-day move, Feb 20, 2026
Okta −9.2%
CrowdStrike −8.0%
Cloudflare −8.0%
Zscaler −5.5%
Fortinet −2.0%
Palo Alto Networks −1.5%

A big security fund, the Global X Cybersecurity ETF (ticker BUG), fell 4.9% and closed at its lowest level since November 2023. A fund (ETF) is a basket of many stocks that you can buy as one.

A security stock fund from February to November 2026

2. Why the fear was only half right

Finding bugs in code is real work, and AI is getting good at it. But it is one small job in security.

Security teams also:

  • stop attacks while they are happening
  • protect logins (who is really signing in?)
  • protect laptops and phones
  • protect networks
  • protect data (who can read it?)

A code scanner does none of that.

Code is one small job in security

3. Then the attackers got AI too: ARTEX

ARTEX is a free tool. Its author built it for penetration testing, which means testing your own defenses the way an attacker would. You give it a target, and AI agents do the rest: they scan, guess passwords, and try to break in. An "AI agent" here is a program that uses an AI model to decide its next step by itself.

The problem: attackers can download it too.

  • October 21, 2026, Microsoft. Microsoft wrote that a group it calls Storm-0940 used ARTEX to automate attacks against "hundreds of organizations globally". In multiple cases, the time from first targeting to privileged access (powerful access) was less than 3 hours.
  • November 6, 2026, Microsoft. A second group, Storm-3358, used ARTEX against banks and other financial companies.
  • November 13, 2026, FBI. The FBI published a FLASH warning about ARTEX and repeated the "under three hours" finding.
  • November 16, 2026, Google. Google's threat team wrote that "the era of autonomous attacks has arrived".

So in one month, Microsoft, the FBI and Google all said the same thing: AI attacks are not a theory anymore.

4. The part for Microsoft 365 admins: "signed in as the app"

This is the part most news summaries skipped.

According to Microsoft, the attackers got hold of the identities of several cloud apps from a vendor called ConnectWise. These are multi-tenant apps. That means one app, made by a vendor, that many customers install in their own Microsoft 365 tenant.

Many companies had already said "yes, I trust this app" in their tenant. So the attackers did not need to phish a user. They signed in as the app:

  • no user
  • no password prompt
  • no MFA (an app does not get an MFA prompt)

The attacker uses the identity of an app the tenant already trusts

The numbers Microsoft published:

  • sign-in attempts on more than 2,250 tenants
  • the majority of the targeting was not successful
  • but in some tenants there was follow-on activity: looking around through Microsoft Graph, adding new credentials to apps, and changing roles
  • Microsoft Defender has an alert for the worst case: "Application self-assigned Global Administrator role". Global Administrator is the top role in Microsoft 365.

Microsoft also saw a password spray from the same infrastructure: more than 151,000 failed sign-ins against 327 accounts in eight organizations, with a peak of about 2,900 attempts per minute. A password spray means trying a few common passwords against many accounts. No human types that fast. That is what "AI gave attackers speed" looks like in a log.

5. So companies are buying more security, not less

News from November 18, 2026.

Palo Alto Networks reported its quarter after the market closed on November 18:

  • revenue: about $3.4 billion ($3.445 billion), up 39% from a year earlier
  • forecast for the full year: raised from $14.10–14.20 billion to $14.52–14.57 billion

That same day, before the report came out, security stocks had already jumped. News reports linked the move to new government warnings about AI-driven attacks.

Stock Nov 18 Nov 19 Nov 20 3 days together
Okta +12.1% +5.7% −7.4% +9.7%
Palo Alto Networks +9.9% +6.4% −5.1% +11.0%
CrowdStrike +8.0% +2.7% −5.8% +4.5%
Zscaler +6.8% +4.0% −5.3% +5.2%
Fortinet +5.9% −0.3% −5.2% 0.0%
Cloudflare +5.7% +0.9% −9.5% −3.5%
S&P 500 fund (SPY) −0.3% −2.0% −3.6% −5.8%

November 18 and November 20, closing moves

One honest note. On Friday, November 20, the whole market dropped, and security stocks gave back part of the jump. A popular summary online said "Palo Alto +20.5% in two days". That number used pre-market prices. On closing prices, the two-day move was +17.0%, and after Friday it was +11.0%. Prices move fast. The lesson here is the threat, not the stock price.

The longer view is clearer. From the close on February 20 (the scare) to the close on November 20:

Feb 20 → Nov 20, 2026
Okta +158%
CrowdStrike +136%
Palo Alto Networks +131%
Security fund (BUG) +76%
Fortinet +56%
Cloudflare +55%
Zscaler +22%
Microsoft +22%
S&P 500 fund (SPY) about 0%

6. What Microsoft says

Microsoft publishes a yearly threat report, the Microsoft Digital Defense Report. The 2026 edition came out on October 1, 2026. Two numbers from it explain this whole story:

  • Attackers now typically turn a newly found bug into a working attack in well below 24 hours (Microsoft's words: "the median time from vulnerability discovery in the wild to weaponization").
  • Companies can need 30 to 60 days to fix critical vulnerabilities that face the internet.

Hours against weeks. That gap is the reason security budgets go up.

What about Microsoft's own security business? Microsoft is not a pure security company, so its stock did not move with the group that week (+0.4%, −2.0%, −2.8%). And Microsoft does not report security revenue every quarter. The last official figure is old: in January 2023, Microsoft said its security business had passed $20 billion a year. Even that old number is bigger than Palo Alto's forecast for this whole year ($14.57 billion at the top end). I could not find a newer official number, so I am not going to guess one.

What Microsoft shipped this month points the same way: defense has to run at machine speed too. At Ignite 2026 (November 17), Microsoft made Microsoft Perception generally available. It runs AI security agents that test, detect and fix. And Microsoft says Defender now stops ransomware attacks in 30 seconds on average. I covered the Ignite security news in my Ignite 2026 recap.

7. Three checks for your tenant (about 10 minutes)

You do not need a new product for this. You need the Microsoft Entra admin center, and Microsoft Sentinel or Defender if you have them.

The three application IDs come from Microsoft's October 21 post. Microsoft's own, longer hunting queries are in that post. The queries below are my short versions. Run them and tune them to your tenant.

Check 1: are the three ConnectWise apps in my tenant?

Search by application ID, not by name. Anyone can give an app a friendly name. The ID cannot be faked.

App name Application (client) ID
ConnectWise Cloud Backup 506e08f9-6e9a-4da4-942b-a099a03644ae
ConnectWise - MS Teams 24205314-2b89-4e07-a2a4-579e898a5640
ConnectWise - MS Graph 0de14c77-b9b0-4137-9360-34878a929827

In the portal: Microsoft Entra admin center → Enterprise applications → All applications → paste an ID into the search box. Remove the "Application type" filter so nothing is hidden.

With a query (Microsoft Defender portal → Advanced hunting):

// 1. Are the three ConnectWise apps in my tenant? (Defender Advanced Hunting)
OAuthAppInfo
| where OAuthAppId in ("506e08f9-6e9a-4da4-942b-a099a03644ae", "24205314-2b89-4e07-a2a4-579e898a5640", "0de14c77-b9b0-4137-9360-34878a929827")
| summarize arg_max(Timestamp, AppName, AppStatus, PrivilegeLevel, IsAdminConsented) by OAuthAppId

No rows is good news: none of the three apps is in your tenant. If you get rows, look at PrivilegeLevel. A "High" app is the one to worry about first.

Check 2: did those apps sign in, and from where?

A service principal is the login an app uses in your tenant. Apps have their own sign-in log, and many admins never open it.

In the portal: Microsoft Entra admin center → Monitoring & health → Sign-in logs → the tab Service principal sign-ins. Set the date to start on October 1 (Microsoft says: review from at least October 15, and from October 1 where you still have the logs). Filter by the application IDs above.

With a query (Microsoft Sentinel → Logs):

// 2. Did those apps sign in, and from where? (Microsoft Sentinel)
AADServicePrincipalSignInLogs
| where TimeGenerated between (datetime(2026-10-01) .. now())
| where AppId in ("506e08f9-6e9a-4da4-942b-a099a03644ae", "24205314-2b89-4e07-a2a4-579e898a5640", "0de14c77-b9b0-4137-9360-34878a929827")
| summarize SignIns = count(), LastSeen = max(TimeGenerated) by ServicePrincipalName, IPAddress, ResultType
| order by LastSeen desc

How to read it: ResultType 0 means the sign-in worked. Look at the IP addresses. Microsoft published these attacker addresses: 144.172.64[.]60, 144.172.64[.]99, 144.172.64[.]128. Microsoft also says: do not stop at the published addresses. Investigate any source you do not know.

Microsoft also listed the user agents it saw (a user agent is the name a program sends when it connects): Python-urllib/3.12, python-requests/2.31.0, python-requests/2.34.2, curl/7.88.1, azure-toolkit/1.0.

Check 3: did an app give out a powerful role?

In the portal: Microsoft Entra admin center → Monitoring & health → Audit logs → filter Activity = Add member to role. Then look at the column Initiated by (actor). A person's name is normal. An app's name needs a closer look.

With a query (Microsoft Sentinel → Logs):

// 3. Did an app (not a person) hand out a role or add a credential? (Microsoft Sentinel)
AuditLogs
| where TimeGenerated between (datetime(2026-10-01) .. now())
| where OperationName in ("Add member to role", "Add service principal credentials", "Add app role assignment to service principal")
| extend ActorApp = tostring(InitiatedBy.app.displayName)
| where isnotempty(ActorApp)
| project TimeGenerated, OperationName, ActorApp, Target = tostring(TargetResources[0].displayName), Result

You will see some rows that are normal. Microsoft's own services (for example Privileged Identity Management) and your HR or provisioning tools also do these things. You are looking for an app that should never hand out roles, or a new credential on an app that nobody on your team added.

Three checks for your Microsoft 365 tenant

If you find something

Microsoft's advice, in short:

  1. Do not delete the app first. Deleting it destroys evidence. Disable it instead (Enterprise applications → the app → Properties → "Enabled for users to sign-in" = No) and keep the affected ConnectWise connections disabled until ConnectWise gives restore instructions.
  2. Keep the logs. Export the Entra sign-in logs and audit logs for the app before they age out.
  3. Look for what was left behind: new credentials on apps, new owners, new role assignments, new consent grants, inbox or forwarding rules.
  4. Then remove what is confirmed bad, and rotate credentials.

8. I checked the viral summary. Here is what held up

This video started from a popular summary that was shared a lot this week. Most of it was right. Some of it was not. So you can trust the numbers above, here is what I found.

Claim in the summary What I found
"CrowdStrike and Zscaler fell about 11%, Okta and Fortinet about 6%" (February) Not the one-day closing numbers. On February 20: CrowdStrike −8.0%, Zscaler −5.5%, Okta −9.2%, Fortinet −2.0%.
"Security fund at its lowest in more than two years" Correct. Lowest close since November 2023.
"Palo Alto revenue up 39% to $3.45 billion, forecast up to $14.57 billion" Correct ($3.445 billion).
"Okta +12.1%, Palo Alto +9.9%, CrowdStrike +8%…" on one day Correct for the close on November 18.
"Palo Alto +20.5% in two days" Used pre-market prices. On closing prices: +17.0%. And on November 20 the group fell with the whole market.
"FBI warned about ARTEX" Correct. FLASH warning dated November 13, 2026.
"Google: the era of AI attacks has begun" Correct. Google's threat report, November 16.
"OpenAI's model found 35 Apple security holes" One Forbes report. Apple has not confirmed it. I left it out of the video.
"Deutsche Bank: 15 million internal files leaked" Real, but it is a leak by whistleblowers to journalists, not a hack.
"Oracle Health: 20 million patients" The breach itself is from early 2025. The new part is the count (nearly 20 million).
"Equinor hit by a cyberattack" Correct (November 17). Equinor says production was not affected.
"Optimizely used to spread malware on sites of 178 organizations" Correct (November 16–17).
"Hackers took control of blood freezers in France" They changed temperature settings at three sites through a supplier. The blood service says no products were lost.

Sources

Official Microsoft sources:

Other sources:

Facts checked: November 21, 2026. Not investment advice. I could not test the three queries against a tenant that has these apps, so treat them as a starting point and tune them to your environment.


Thanks for reading! I'm Cuong, and on CD Security Lab I rebuild real Microsoft security cases in a safe lab and explain them simply. Follow along:

Got a question about your own tenant? Drop it in the comments.

來源:DEV Community · dev.to