Cyber Weekly: Your MSP was the target

Two management platforms that small businesses never chose, and never see, were the story of the week. One pushed attacker commands to 288 machines in 90 minutes. The other had 33 working exploits dumped on it at once, and 19 of the servers that fell were already patched. That is the shape of the problem now. Ask a 30-person company who handles its IT and the answer is a name, not a product. So when the tools behind that name break, the business does not experience a vendor incident. It experiences its own outage, its own data leak, its own locked files. The rest of the week rhymes. A law firm lost Microsoft Copilot tokens over a screen share. Phishing walked past email filters 41.3% of the time. The FBI said rented AI reaches privileged access in under three hours. Almost none of it needed an unknown bug.

Start with the invoice nobody thinks about. Your IT provider pays for a remote management tool, configures it once, and forgets it. Years pass. Extensions get added by whoever needed something that week. None of it shows up in your quarterly review. That is the quiet infrastructure behind your patched laptops and your nightly backups, and this week two of those platforms handed attackers the keys.

288 endpoints in 90 minutes. 33 zero-days in one management platform. 21 servers rooted, 19 of them already patched. 41.3% of phishing never reaches your filter. 5,000 developers exposed for 11 hours. 642 ransomware victims in eight days.

This week in cybersecurity news: eleven stories, one pattern

ScreenConnect: the extension was the door, not the CVE

An intruder got into cloud-hosted ScreenConnect panels and used them to push commands at managed endpoints. Huntress counted 288 hosts across 48 organizations in roughly 90 minutes, with about a third reached inside the first ten. Six MSPs were involved.

The way in was not a zero-day. A third-party extension called RMM+ (versions 1.0.31 and 1.0.32) authenticated with a shared access key that defaulted to blank when nobody set one. That quietly skipped ScreenConnect's own accounts, roles, and MFA. ConnectWise disabled five RMM+ family extensions across roughly 400 partner environments.

The payload, which Huntress calls SCKit, installs rogue ScreenConnect clients that reuse the real client's own DLLs, so the machine still looks normal. It also carries code that waits for a technician to connect, infects the technician's workstation, and spreads to other panels.

One correction, because the detail is circulating wrong: CVE-2026-84869 was patched back in September, and Blackpoint's testing found exploitation against patched clients failed. This campaign abused an extension, not that vulnerability. CyberScoop has the details.

If you outsource IT, ask which extensions are installed on their ScreenConnect server and whether every one has a passphrase set. That question takes a minute, and it is the whole story.

The endpoints were fine. The console that managed them was the way in.

N-able took its cloud offline after 33 zero-days landed at once

Independent researcher Sean Bales published 33 vulnerabilities in N-able N-central with working proof-of-concept exploits, collectively tracked as Able2Shell. Credential disclosure, arbitrary file write, directory traversal, XSS, XXE, and remote code execution as root.

N-able took its Hosted N-central environment offline, told on-prem customers to shut down servers, and shipped hotfix 2026.4 HF2. Then a second set, 4EverAble2Shell and Agent2Shell, landed on Thursday.

Attackers chained an Apache Axis flaw (CVE-2026-90436, CVSS 9.8) to root 21 N-central servers across seven countries. 19 of those 21 were already running Hotfix 2, which is the sentence to sit with. The payload was XMRig, configured to use only half the CPU cores so nobody noticed. N-able has since disabled the legacy API by default and switched off the Intune integration. The Hacker News tracked the timeline.

N-central is how a lot of small businesses get patched, backed up, and monitored. When the console is compromised, the attacker does not need to breach thirty clients one at a time. They log in and inherit all thirty.


Luna Moth stole Copilot tokens during a screen share

The extortion crew that has been going after law firms since 2023 changed tactics. No remote management tools, no manual exfiltration. A blank-subject email with a deliberately broken PDF, a follow-up phone call from "IT support" offering to fix it, a video call, and a request to share the screen.

Then the helpful technician walks the victim through opening Microsoft Copilot in a browser, opening developer tools, finding the request Copilot sends to Microsoft Graph, and copying the application access token into a public paste site. From there the attacker queries Graph directly and pulls files. Dark Reading has the write-up.

ReliaQuest believes the firms being retargeted are the ones that refused the first demand, which is a grim sort of loyalty.

Three controls actually help here. Block browser developer tools on managed machines. Restrict screen sharing to people with a business reason for it. And turn on Continuous Access Evaluation so a stolen token dies faster, because as we wrote recently about MFA, the prompt never fires when the attacker holds a token instead of a password.


The FBI says attackers are renting AI, and it works in under three hours

A FLASH advisory says multiple threat actors have been weaponizing ARTEX, an autonomous penetration testing framework written in Go, since at least August 2026. The line to underline: intrusions progressing from initial targeting to privileged access in under three hours, run by people with minimal technical expertise, against several targets at once.

Google's threat intelligence group followed by saying it expects fully autonomous agentic attacks within six months, and named four frameworks in active use: ARTEX, CAIRN, CyberStrikeAI, and HexStrike AI. ARTEX has also been linked to the October ConnectWise Cloud Backup supply chain breach, which ties this story back to the first one. BleepingComputer covered the advisory.

For a small business, the assumption that used to hold was that nobody would bother writing a custom exploit for your forgotten portal. That assumption is finished. Rate-limit your login endpoints, take admin interfaces off the internet, and start paying attention to bursts of automated requests that used to look like background noise.

Writing an exploit used to be a skill test. Now it is a subscription.

41% of phishing never touches your email filter

KnowBe4's latest threat report, released this week, put a number on something most of us suspected. Phishing volume is up 29.8% year over year, roughly one dangerous email per 733 messages, and 41.3% of it never reaches a secure email gateway at all.

The mix has shifted hard toward targeted work. Spear phishing up 41.1%, pre-attack reconnaissance up 33.7%, company impersonation up 28%, while spray-and-pray categories collapsed.

Three kits account for 90.2% of kit-driven phishing (Sneaky 2FA at 41.4%, FlowerStorm at 32.6%, ONNX Store at 16.2%), and all three use reverse-proxy adversary-in-the-middle to steal live session cookies. Business email compromise remains the largest category at 51.2%.

Which means a password reset no longer ends the incident. Rotate sessions, not just passwords. The timing matters too: 90.4% of attacks land Monday through Friday, with Monday the heaviest at 20.6%.

One caveat: the report carries an October cover date, so treat this as published this week rather than broken this week. The Journal has the summary, and our piece on why awareness training alone is no longer enough is the companion read.


Cursor shipped a malicious Claude Code extension for 11 hours

Cursor confirmed a breach that let attackers distribute malicious versions of Anthropic's Claude Code extension (2.1.312 and 2.1.313) through its marketplace on October 29. Up to 5,000 developers were exposed, and the tampered builds stayed live for about 11 hours.

The root cause was a credential stolen from AI infrastructure provider Fireworks AI. That led to internal systems and historical coding-agent traces containing unredacted credentials, which the attacker then used to go further.

The extension ran JavaScript that installed persistent SSH backdoors, though the coding was sloppy enough that the backdoor likely failed on many Windows and macOS machines where SSH was not already enabled.

Two cheap lessons. AI coding tools are now a supply chain with the same risk profile as npm, which is the pattern behind the GitHub Actions theft we covered. And storing unredacted credentials in logs and agent traces is how one breach quietly becomes three. Notion confirmed customer data exposure through the same Fireworks AI incident, affecting under 1% of workspaces.


Check Point gateways, and one wave that ended in a wiper

Beginning October 30, attackers scanned for unpatched Check Point Security Gateways and hit them through Mobile Access portals using PHP2Shell (CVE-2026-109618 and CVE-2026-109619), roughly a week after patches shipped. A first wave against Israeli organizations ended in destructive activity. A broader second wave on November 2 hit multiple countries without destruction.

From the portal service account they escalated to root through CVE-2026-123039 in the DynamicCLI IPC interface, dropped PHP webshells plus the TSGA backdoor, pivoted over SSH to a VMware ESXi host, and left a wiper with a politically themed note. CISA added CVE-2026-123039 to KEV on November 13 with a November 16 deadline. Check Point Research published the exploitation analysis.

A firewall is the one security product nobody expects to be the way in. If you run Check Point with Mobile Access enabled and you are not on the current build, that is a same-day job, followed by a webshell check.


Krebs named the man behind Qilin, and the volume explains itself

Brian Krebs unmasked the administrator of Qilin, now the most active ransomware affiliate program in the world. At least 365 claimed victims in three months, affiliates keeping 80% or more of each payment, and a 36-year-old man in St. Petersburg traced through a decade of operational security mistakes.

The affiliate model explains the volume. The person encrypting your server is a freelancer who bought a toolkit, so the threat model is volume rather than targeting. Manufacturing was the hardest hit sector in November at 77 victims (12%), just ahead of healthcare at 64. Qilin also claimed the University of Texas at Arlington, with 41,000 students.

The month so far: 642 ransomware victims through November 16, a seven-day average of 47.9, and a single busiest day of 97 on November 9. If you have never put a number on an hour of downtime, do that before the invoice arrives unplanned. Krebs has the full investigation.


TSMC's employee data leaked out of somebody else's cloud

TSMC filed with the SEC confirming that a cloud service provider was breached and employee data exposed. The company says its own systems were untouched and the data is unrelated to operations or customers, but it verified the leaked records matched data held by the unnamed vendor.

The trigger was a November 8 forum post offering 6.9 GB of employee data, including contact details, identification documents, and banking information, for 3 BTC, roughly $200,000. The Record reported the filing.

Substitute your own vendors. Your HR platform, your payroll bureau, your benefits administrator. A semiconductor giant can absorb this. A 40-person company whose payroll provider gets hit cannot. Ask what employee data each vendor holds and whether it is encrypted at rest, which is the first page of any honest third-party risk conversation.


Japan's vendor cascade: 77 million emails and a government bailout

A November 2 attack on Japanese email distribution provider Arara exposed 77 million email addresses across nearly 400 customers, including TKC, Sega, and the city of Osaka. Osaka confirmed more than 1 million addresses tied to its administrative services platform had leaked, covering anyone emailed between July and October 31, and warned 2.8 million residents about phishing.

Then the attackers used Arara's own infrastructure to send millions of malicious emails impersonating credit card companies. Japan's government responded with a ¥10 billion package, about $63 million, covering AI-assisted defense, public-private responders, and expanded active cyber defense powers.

This is the vendor cascade pattern in its purest form, and small businesses were the ones locked out of portals and hit with the follow-on phishing.

Japan has logged more than 600 publicly reported incidents in the first nine months of 2026, already past all of 2025, with roughly 380 million personal records taken this year. The money also went to victim recovery, not just defense. The Record has the Osaka details.


Office IP cameras are being turned into crypto miners

Huntress documented attackers exploiting unpatched GeoVision camera software to run XMRig miners across 13 organizations, with most activity between November 8 and 10. Everything traced back to GeoStartupAgentService.exe, frequently running as NT AUTHORITY\SYSTEM.

The chain is nastier than the payload suggests. Bring your own vulnerable driver (wsftprm.sys) to tamper with Microsoft Defender, plus a PowerShell script that drops a Go loader using BitTorrent DHT and PKARR to find its command server, so infrastructure can move without losing access.

Cheap IP cameras on the office network are the classic forgotten asset. They run as SYSTEM, they ship with firmware from 2019, and nobody patches them. Put them on their own VLAN, or accept that a miner is using your bandwidth and your electricity.


Also worth a minute

  • UnrealIRCd CVE-2026-114713 (CVSS 8.6) went into KEV on November 13, exploited one day after the hotfix shipped. The maintainer called it the worst issue in 15 years. Fixed in 6.2.7.1. Only matters if you run a community or support IRC network, but the shrinking patch-to-exploit gap is the real lesson.
  • Red Hat Quay CVE-2026-123710 (CVSS 9.8) allows unauthenticated token-signing takeover, and the attacker's signing key survives restarts. Relevant if your developers run a container registry.
  • BYOD ransomware is having a moment. Vanderbilt University has more than 1 million SSNs claimed against it, and Texas HHSC has 2 million Texans affected.
  • Maersk faces a $5 million demand from WhiteArmor, with about 100GB claimed and 33 super-admin accounts in plaintext, taken from an inherited legacy platform. Best acquisition-debt cautionary tale of the week.

Three questions for whoever handles your IT

None of this week's damage required a brilliant attacker. It required a blank passphrase, an extension nobody reviewed, a token in a paste site, and a monthly invoice that made everyone assume someone else was watching.

So ask three questions this week. Which management and remote access tools can reach our machines, and who reviews their extensions? What happens to a stolen session token, and how fast can we kill it? Which vendor holds our employee and customer data, and what is their notification commitment in writing?

If the answers are vague, that is the finding. Vagueness about access is how 288 endpoints get touched in 90 minutes.

 

Book a 15-minute security check
We will map the tools and third-party access that reach your machines, no obligation.