Cyber Weekly: The credentials were valid

The attacker had a valid login. A stolen app secret skipped MFA, a host leaked its customers' passwords, and a signed app shipped malware.

Eleven stories made the cut this week, which is more than usual for early November. Almost none of them needed a clever exploit. One caveat before the numbers, because it matters: exposure figures are not confirmed misuse, and early vendor estimates tend to move.

27 organizations reached through stolen app secrets. 10,000,000+ sites running the affected contact form plugin. 7,700,000 records exposed at a retailer. 241 ransomware victims in seven days. 37 of 50 machines encrypted at one school. 6 days a vendor stayed silent.

Cybersecurity news this week: eleven stories to check

Contact Form 7 lets an anonymous visitor leave a file on 10 million sites

Contact Form 7 versions 6.2 through 6.2.3 fail to check the type or extension of files uploaded through a form field, and a cleanup bug leaves the file sitting on disk. An unauthenticated visitor can submit a form and place a file of their choosing on the server. On stock Debian or Ubuntu Apache, where directory listing is on and .phar is mapped to PHP, that file runs. On nginx, or Apache with AllowOverride All, it is "only" an arbitrary file write into wp-content/uploads, which can still be served as HTML or SVG from your own domain.

WPScan rates it CVSS 7.5 and the CISA ADP score is 9.8, which tells you something useful: exploitability depends entirely on server configuration, and most businesses have no idea which side of that line their host sits on. The fix landed on October 29 in 6.2.4, and the current release is 6.2.5. Only forms with a [file] field are exposed. A public proof of concept is scheduled for November 19. See the GitHub advisory.

This is the second WordPress problem in as many weeks, after the core update flaw we covered earlier. If you run WordPress with this plugin, check the version today, then check whether any form on the site accepts a file. A week of quiet is not much, and "our host is probably fine" is not the same as knowing.


One operator, 27 organizations, and no MFA prompt to reject

A single actor signed into Microsoft 365 and Azure tenants as the victims' own applications, using valid stolen client secrets for Entra ID app registrations. Researchers at Todyl counted 45 targeted registrations, more than 33 successful authentications, and 27 organizations reached, including MSP clients in government, energy, food and agriculture, transportation, construction, legal and education. Microsoft corroborated the mechanics two days later.

The scale numbers are the part worth reading twice: 81 files, roughly 2,279 mailbox items, bursts of 24 sign-ins in 10 seconds, and 95% scripted logins from a single IP address. What was taken is the material that unlocks the next stage: password lists, private keys, VPN profiles, MFA and RMM API keys, backup storage credentials, firewall and RDP configuration. Four AI service tokens were issued at two MSPs.

The app was valid. The secret was valid. The MFA prompt never appeared.

If your business uses an MSP, this is the story to forward to them. One compromised app registration can expose every client it manages, and the usual small-business controls do not apply, because user-scoped Conditional Access has nothing to evaluate when the sign-in is an application. Ask when app secrets were last rotated, whether workload identities use certificates instead, and whether anyone reviews sign-ins from unfamiliar infrastructure. Secrets left in code tend to stay valid for years, so rotation beats hunting. Todyl's write-up has the technical detail.


SonicWall confirms MFA can be skipped with only a password

An authentication enforcement flaw in SonicOS SSLVPN lets someone holding credentials for an alternate authentication method bypass SAML, and with it any multi-factor authentication riding on SAML. In practice, a VPN login that should have required an Okta or Entra prompt succeeds with a password alone. The vendor score is CVSS 7.1, which is the trap: it reads as skippable for a flaw that erases your second factor.

Gen7 devices up to 7.3.3-7015 and Gen8 up to 8.2.2-8015 are affected, fixed in 7.3.4-7013 and 8.2.3-8015. CISA added it to the Known Exploited Vulnerabilities catalog the same day with a November 20 federal deadline, and SonicWall confirmed it investigated exploitation. There are six other CVEs in the same advisory, so partial patching is not a plan.

One detection idea costs nothing: compare SSL-VPN login records against your identity provider's sign-in log and look for sessions that succeeded without a matching MFA event. We wrote about why this class of bypass keeps working in our piece on MFA and phishing kits.


DreamHost took ten weeks to notice, and an outage is what noticed

An intruder exploited a flaw in the third-party software behind DreamHost's browser-based File Manager and collected the FTP and SFTP credentials of everyone who logged in between August 15 and October 23. They also reached webmail and took session tokens, and in a small number of cases may have read mailbox contents. The company says "well under 10%" of accounts were affected, against roughly 400,000 paying customers. A Washington state filing lists 871 residents.

The uncomfortable part is the timeline. Nobody noticed the intrusion until an outage forced the issue. And a stolen webmail session token is not just a credential problem, it is a business email compromise waiting to happen: the attacker is inside the mailbox, reading supplier threads and watching for invoices. TechCrunch has the details.

If you logged into that file manager during the window, rotate the FTP and SFTP password even if it was unique, and change it anywhere it was reused. Then check the mailbox for forwarding rules and connected apps you did not add.


Fake OpenAI and SpaceX invitations are farming work accounts

Push Security documented a campaign using invitations to OpenAI livestreams, an invite-only SpaceX "Starship Product Briefing", and an Anthropic session at re:Invent, each promising a reward for signing in with a work account.

Choosing "Continue with Microsoft" runs a device code phishing flow through a phishing kit: you are shown a code and sent to Microsoft's real sign-in page to enter it, which passes every check a careful user has been taught to make. Choosing "Continue with Google" drops you into an adversary-in-the-middle proxy that captures the password, the MFA response, and the session cookie.

The AI branding is doing the work, because your staff genuinely want these tools. The related kit logged 618 sandbox analyses between June and October 21, growing from about 20 a month in June to 200 in September, across 173 hosting domains. We have written before about why awareness training alone stopped being enough for AI phishing.

Three fixes, none expensive. Block device code flow in Conditional Access by default, and allow it only for meeting room hardware. Move admins and finance to phishing-resistant MFA with passkeys or FIDO2 keys. And give staff one rule that survives a convincing page: never type a code into a Microsoft or Google sign-in page unless you started the process on the device in front of you.


A school lost 37 of 50 machines to what looks like an automated loop

A 3AM affiliate encrypted 37 of nearly 50 reachable machines from a single unmanaged host, according to Bitdefender's reconstruction. The evidence for automation is behavioural: repeated commands, slightly modified on each attempt, issued at intervals of roughly 40 to 100 seconds. The honest framing is "consistent with an agentic loop", not "AI hacked a school", since the researchers never saw the attacker's machine.

The rest of the timeline is a checklist of familiar gaps. A modified PowerShell tool pulled credentials from three Veeam backup servers, and a Chrome credential database was copied. The attacker also created two backdoor domain admin accounts, configured two-way trusts between three separate Active Directory domains, and then ran a Rust encryptor. Activity spanned September 10 to 19. The one control that held was antivirus, which the attacker tried to disable and could not. BleepingComputer has the reconstruction.

For a small organisation, the lesson is not "buy an AI defence". It is that one unmanaged machine was enough of a door, that backups were reachable from the domain they were meant to protect, and that nobody had reviewed the domain trusts since the last acquisition. Our piece on why backups and recovery are two different projects covers the second point.


IBM Aspera Faspex: an ordinary login becomes full admin

An incorrect authorization flaw in Faspex 5.0.0 through 5.0.19 lets any authorized user craft a message that grants full administrative privileges. IBM published CVSS 9.8 and no workaround. CISA added it to KEV on November 9 with a November 12 deadline for federal agencies, which is today. The fix is 5.0.19 PL1 on Linux.

Faspex is the tool design studios, engineering firms, and media or life sciences teams use to push very large files to clients. Those servers also hand logins to outside clients and freelancers so they can send packages, and every one of those accounts is now a path to admin. An admin can read every package on the server and reroute deliveries. Its EPSS score of 0.004 shows how far exploit prediction lags a KEV listing.

Treat this as still developing, since the bulletin landed on the last day of the window. If you receive or send files through a client's Faspex instance, you sit inside the blast radius of their patch discipline. IBM's bulletin has the affected versions and the fix pack.


Code signing held. The attacker was inside the pipeline.

Attackers compromised two Ethersphere GitHub accounts, one a maintainer with force-push rights to an unprotected branch and one the release bot, then published @ethersphere/bee-js 13.3.0 to npm using the project's own automation. The release carried valid SLSA provenance. A trojanized Swarm Desktop 0.56.2 shipped signed with the project's Developer ID and notarized by Apple.

The payload steals the Swarm node's wallet key and the plaintext password from its config file, then installs a backdoor: a Go remote shell on Linux and macOS, a hidden ScreenConnect client running as SYSTEM on Windows. The npm release sat tagged as latest for about 75 minutes. The desktop release stayed up for roughly 8.5 hours and drew 65 downloads from GitHub. Clean versions are 13.2.0 and 0.56.3. Copies survived removal on a third-party npm mirror and in GitHub source archives by commit SHA.

Provenance tells you who built an artifact. It never tells you what the artifact does.

You probably do not run Swarm, so take the general lesson instead. If your team's dependency policy is "signed packages from reputable publishers", the signature chain here was unbroken straight through the attacker's commit. SafeDep has the technical write-up.


WatchGuard's MFA gateway has critical flaws and does not update itself

Four vulnerabilities in the AuthPoint Authentication Gateway, the on-prem Windows component bridging WatchGuard's cloud MFA to RADIUS clients, the ADFS agent, and Active Directory or LDAP. The worst needs only network access to the gateway host, no login at all, to disrupt authentication and delete local directory sync settings. The others allow an on-path attacker to modify traffic, a replay of a captured request to perform admin actions, and a forged authentication result through the LDAP callback. Three of the four carry critical ratings, between CVSS 8.9 and 9.2.

Here is the detail that should annoy you: the gateway does not automatically upgrade, and WatchGuard's own documentation says so. A business that checks the cloud console, sees the latest version, and assumes it is current could be running a build from 2022. Affected versions are 1.0 through 7.6.0. WatchGuard says it is not aware of exploitation and none of the four are in KEV, so this is a get-ahead story rather than a fire.

Two things today: confirm the version actually installed on the gateway host rather than the console, and make sure TCP 9000 to 9003 are not reachable from the internet. One of the four flaws needs nothing more than that.


Your vendors hold your customer list, and some of them tell you late

Iris Plaza, the e-commerce arm of Iris Ohyama, found a suspicious file in an external cloud service on October 28 and took its storefront offline. The November 5 disclosure confirmed that roughly 7.7 million records may have been viewed or taken: 5.8 million members and 1.9 million non-members, with names, addresses, phone numbers, emails, gender, dates of birth, and encrypted passwords, but no card data. The vulnerability was not in the company's own network. It was in a service it did not operate, holding a copy of its customer list.

Then there is Who Gives A Crap, the most relatable breach of the week. A US marketing software provider told the brand about an intrusion on November 6. The access began on October 30, the provider identified it the next day, and declared it "secured" on November 1. That is six days of silence before the customer was told, and the customer then had 24 hours to notify its own customers. Exposed fields: names, emails, phones, mailing addresses, and the total amount each person had spent. The provider has never been named.

Your incident response clock is not your vendor's clock.

So ask a boring question this month: which of your vendors holds names, addresses, and spending history, and do they need all of it? Our guide to third-party risk is a reasonable place to start the inventory.


The ransomware bill is downtime, not decryption

Queens Public Library's network went down on November 4 across 66 branches serving 2.3 million residents: website, mobile app, public computers, checkout and returns, Wi-Fi, and digital media. Six days later, Wi-Fi and checkout were back and the website was still offline with no timeline. No ransom demand, no claim, and nobody has said whether user data was taken. The Record reported it.

In Canada, the Burnaby School District's stolen database turned up on Qilin's leak site. Screenshots show folders labelled "Harassment Investigations", "Discipline", "Payroll", and "Mental Health and Wellness", with roughly 1 TB claimed. Attackers got in during September, the listing appeared in October, and notices to former staff went out on November 5, alongside 24 months of credit monitoring.

Week-wide, researchers counted 241 ransomware victims across 44 countries claimed by 48 operators, and one tracker put October at 953 victims, up 21% on September, with 117 in healthcare. A library that cannot serve a website for six days is a small business that cannot take orders. The real invoice is the unplanned week, not the decryption fee. If you have never put a number on an hour of downtime, this is a good place to start.


Also worth a minute

  • Microsoft published a CVSS 10.0 and told you to do nothing. CVE-2026-109469 in Entra Device Registration was fully mitigated server-side before the record went live. No patch, no action. A useful reminder that a score describes severity, not your to-do list.
  • Google dismantled a botnet inside its own certified hardware. The XCPE backdoor shipped in firmware on more than 1.3 million Android TV boxes. Google took legal action and revoked the manufacturer's certification, but the malware is still on the devices.
  • npm had another bad week. Twelve malicious releases across ten hijacked packages, with the loader fetching its command address from a public Ethereum transaction so the attacker can move infrastructure without republishing.
  • The AI vulnerability pipeline went industrial. One agentic pipeline reportedly reproduced 1,724 Linux kernel bugs in 30 hours and generated 948 working local privilege escalation exploits at roughly $1,279 each. Your patching cadence was designed for a slower attacker.

Start with the logins you already trust

The thread running through this week is not sophistication. It is valid access: a secret nobody rotated, a VPN login that skipped a prompt, a signature that certified the wrong thing, a vendor that held the customer list. None of that gets fixed by buying another dashboard.

The work that helps is unglamorous. Inventory the app registrations, service accounts, and vendor integrations that can reach your data. Rotate the secrets behind them. Confirm your MFA cannot be bypassed by design. And make sure someone knows what to do when a supplier calls, because the call usually comes late.

If you are not sure where to look first, start with whatever holds customer records or connects directly to production. That is usually where the quiet access lives.


Book a 15-minute security check
We will map the credentials and third-party access that matter most, with no obligation.