A working examiner's read of the week's most interesting compromises. What happened, how it happened and what it means for anyone who has to reconstruct the event afterward.
This edition covers November 15 to 20. One sentence from Barracuda sums it up. The company gave a cloud access key to an AI vendor for a project and never retired it when the work ended. Its own verdict: "That is on us." The same failure runs through most of the week. A data platform's access stayed in customer accounts after the customers stopped using it. A ScreenConnect add-on was "usually added for a project and left in place," in Huntress's words. A security company says it was breached through a vendor. Sources say that vendor moves SharePoint files.
None of these attackers had to break anything. Each used access somebody had granted on purpose and then forgot. That is good news for the investigation in one way. A credential that sat idle for months and then woke up is one of the cleanest signals in any log. It only helps if the logs from the quiet months still exist.
Barracuda: a vendor key nobody retired
On November 16 Barracuda described an intrusion it calls "machine-speed." BleepingComputer reports the details. The attacker used a stolen cloud access key and ARTEX, an open-source AI penetration testing framework. Barracuda had issued the key to a third-party AI vendor for a past project. It was never retired. Barracuda confirmed that the key sat in a configuration file in the vendor's GitHub repository. It believes the attacker took it from there in October.
Between November 1 and 2 the attacker worked through about 10 Barracuda cloud accounts in 44 hours. Barracuda found the activity on November 4. The attacker moved from system to system with credentials found along the way. It reached a data analytics environment, source-code repositories and Kubernetes clusters. It ran commands looking for cryptocurrency secrets: private keys, Ethereum keystore files and seed phrases. It tried to persist by deploying a container that mounted the host's file system. It tried to create a new cloud administrator identity and was blocked. It also reached a system that stores customer integration credentials for Barracuda XDR. Barracuda says it found no evidence that those credentials were accessed. It has still asked potentially affected Managed XDR Cloud and Monitored Endpoint customers to revoke and recreate their integration credentials.
The ARTEX link rests on pattern matching. Barracuda says the machine-style labels, probing methods and built-in hunting routines matched the framework. It did not recover a binary. BleepingComputer notes the case closely resembles a recent breach at the cloud host Railway.
For an examiner, "machine speed" changes the clock but not the method. Forty-four hours across ten accounts is too fast for a person to read along with. The question to answer first is the key's own history: when it was created, who it was for, when it was last used legitimately and when it was first used by someone else. Cloud audit logs tie every call to the access key that made it. A key that went quiet when a project ended and came back to life months later draws its own timeline. The failed calls matter as much as the successful ones. Cloud audit logs generally record a denied request too. The blocked attempt to create an administrator is written evidence of what the attacker wanted next.
Qubole: the same failure across many customers
Barracuda's case is one key. The Qubole campaign shows the same gap at scale. On November 9 Wiz published research with Sentra and Mandiant on an attacker who used IAM credentials for an AWS account associated with Qubole, a big-data platform. Between October 14 and 26 the attacker used them to assume the integration roles that Qubole's customers had created in their own AWS accounts. Wiz says every one of those role requests carried the correct External ID and every one succeeded. How the attacker obtained Qubole's credentials is not known.
Many of the roles were dormant. Sentra puts it plainly: "Qubole roles were still in place long after companies had stopped using Qubole." Once inside, the attacker read secrets from AWS Secrets Manager, SSM Parameter Store and Lambda environment variables. Sentra counts about 100 passwords and keys read in 7 minutes. Wiz says the attacker tested what it found against AWS, Entra ID and Okta. It read whatever S3 buckets each role could reach. Sentra counts zero write actions. Nothing was changed or deleted.
On November 11 Wiz updated its post to note that an actor calling itself n0n had claimed responsibility. On November 12 Expedia confirmed to Skift that attackers reached its systems through a compromised third-party data platform provider. Security Affairs reports Expedia's statement and n0n's claim of more than 830 million loyalty program records from 35 programs. Expedia has not confirmed those figures. It has not named its provider either. Several outlets connect Expedia to the Qubole campaign. Treat that link as reported, not confirmed.
Wiz's post is unusually useful to an investigator. The attacker's sessions used role session names beginning quiam-. Wiz recommends reviewing CloudTrail from October 1, 2026 in every account that holds an integration role. It warns that the External ID is logged in plaintext in CloudTrail, so it should be treated as non-secret. It also warns that "Rotating a credential does not end temporary sessions that were already issued."
Two evidence points deserve emphasis. First, CloudTrail management events do not record S3 object reads. Without S3 data events or server access logs, the list of files taken cannot be rebuilt from CloudTrail alone. Sentra adds that files using default encryption "leave no trail." Second, retention is now the deadline. CloudTrail's built-in event history keeps 90 days of management events. Activity from October 14 starts to fall out of it in mid-January. Any account without a trail writing to S3 should export what it has now.
Infoblox: breached through the migration vendor
On November 16 TechCrunch reported that Infoblox is investigating "a security event involving unauthorized access to certain company information" through a third-party vendor. Infoblox says it "immediately activated our incident response procedures" and brought in outside experts. It says it has "no indication that any Infoblox products" were involved. It did not say whether cloud services or build and release processes were affected. It would not say how many customers were affected or whether it had been contacted by the attackers.
Two people with knowledge of the incident told TechCrunch that the vendor provides SharePoint data migration services. One said sensitive customer data was taken. Infoblox has confirmed neither point. Nobody has publicly claimed the attack.
A migration vendor is a particular kind of risk. Its job is to copy everything, so it usually works through an account or application with broad read access across the tenant. Its activity looks like bulk reading because bulk reading is the job. Separating the vendor's legitimate copying from an intruder's depends on detail: which identity read which files, from which address and at what time against the migration schedule. In Microsoft 365 that detail sits in the unified audit log. The default retention for Audit (Standard) is 180 days, so the window is finite here too.
SCKit: ScreenConnect's own Run command turned on its customers
Our last edition did not cover this one. It broke on November 11 and it has kept developing. On November 11 at about 17:47 UTC Huntress saw suspicious activity across several ScreenConnect environments. Within about 90 minutes attacker-controlled ScreenConnect clients were installed on at least 288 endpoints at 48 organizations. About a third were reached in the first 10 minutes. ConnectWise published an advisory later that day and disabled five third-party extensions across about 400 partner environments. It says it found no evidence of a compromise of its own infrastructure or of a vulnerability in ScreenConnect itself. A second wave on November 13 hit another 205 hosts at 41 organizations from new infrastructure.
Huntress traced the access to RMM+, a third-party extension. Versions 1.0.31 and 1.0.32 used a shared access key that defaulted to blank. In Huntress's words, "if an administrator had never configured the extension's access key, RMM+ would accept a blank one." The extension bypasses ScreenConnect's user accounts, roles and MFA. The attacker used the panel's Run Command feature to push commands to groups of endpoints as SYSTEM. On November 16 heise reported that on-premises customers have to disable the extensions themselves. ConnectWise's instruction is blunt: "Do not wait for an alert or evidence of compromise before taking this action."
Huntress calls the toolkit SCKit. It downloads an installer to C:\Windows\TEMP\sc.msi and installs a rogue client. A scheduled task named TrapWatch runs as SYSTEM every 10 seconds. It stops the legitimate ScreenConnect client and takes its connection slot. The replacement reuses the machine's real client libraries, so on the relay the host still looks normal. SCKit also adds a Run key value named RelayTrapLiteAlert, adds Defender exclusions, edits the hosts file and sets a netsh interface portproxy rule. A second stage tries to steal a technician's ScreenConnect session cookie and replay it against the panel. Huntress reports that "the console-takeover attempt failed in every recovered host log we have checked so far."
For the examiner the first rule is Huntress's: "Do not triage infected hosts over ScreenConnect." On the panel side, the Run Command and Toolbox history show what was pushed and when. Huntress's key observation was commands arriving with no matching login. A Run Command entry with no session behind it is the signature. On each endpoint, look for a second ScreenConnect client service, the TrapWatch task, the Run key value and the RelayTrap folder under C:\ProgramData. Windows records a new service in the System event log as event ID 7045. It records a new scheduled task as event ID 4698 if object access auditing is on. Portproxy rules are stored in the registry, so they survive a reboot and an image will show them. Huntress also warns that "isolation alone does not remove the persistence."
LiteLLM CVE-2026-40217: the AI gateway as a root shell
On November 18 CISA added four flaws to its Known Exploited Vulnerabilities catalog. The one most likely to surprise people is in LiteLLM, the open-source proxy many companies put in front of their AI model providers. CVE-2026-40217 sits in the POST /guardrails/test_custom_code endpoint. According to the GitHub advisory, it runs user-supplied Python in a hand-rolled sandbox that can be escaped with bytecode-level techniques. The code then runs in the proxy process. In the default Docker image that process runs as root.
The advisory says the endpoint needs a proxy-admin credential in default configurations. Affected versions run from 1.81.8 up to 1.83.10. The advisory gives the fix as 1.83.10 in one field and 1.83.11 in another. Upgrade to 1.83.11 or later. If that cannot happen today, the advisory suggests blocking the endpoint at a reverse proxy or API gateway. The flaw was published in May. Its arrival in the KEV catalog in November means CISA now has evidence of real attacks.
The forensic point is what a gateway like this holds. It exists to keep the API keys for the model providers behind it. Barracuda's intruder went hunting for keys. An AI gateway is a key store with a web front end. If a LiteLLM instance was exposed, check the reverse proxy or access logs for requests to the guardrail test endpoint. Legitimate use of it is rare. Then rotate the proxy's own master key and every provider key it held. Preserve the container before redeploying. Restarting a container from its image throws away its writable layer, which is where an intruder's files would be.
The other three additions on November 18 were Synology DiskStation Manager CVE-2024-10441, Apache Axis CVE-2026-90436 and REDCap CVE-2026-90817. More on them in the watchlist.
NVIDIA: a claim with a countdown
On November 19 BleepingComputer reported that the Gentlemen ransomware group had added NVIDIA to its leak site. The countdown ends on November 28. The group has not shared any proof. NVIDIA says it is investigating "an incident that may have affected a third-party partner." It adds: "We have found no indication that NVIDIA systems or networks were affected."
That statement puts the evidence somewhere NVIDIA does not control. If data was taken from a partner, the partner's logs decide what was taken. NVIDIA's own logs can only show what the partner was given. When the leak or the proof arrives, the useful test is the same as with any extortion claim. Compare each claimed file against what the partner actually held and when it was shared.
On the watchlist
- Synology DSM CVE-2024-10441. CISA added it on November 18 and describes it as improper encoding or escaping of output. It is a 2024 CVE. Its appearance in the exploited catalog two years on is a reminder that NAS firmware stays in service long after anyone last looked at it.
- Apache Axis CVE-2026-90436 and REDCap CVE-2026-90817. CISA added both on November 18. It lists the Axis flaw as unsafe reflection and the REDCap flaw as code injection. REDCap is the research data capture platform from Vanderbilt University. Institutions running it should treat this as urgent.
- Check Point CVE-2026-123039, now with a wiper attached. Last edition listed this flaw from CISA's November 13 batch. BleepingComputer's report on Check Point Research's findings fills in the chain. Attackers reached unpatched gateways through the Mobile Access portal using the PHP2Shell flaws (CVE-2026-109618 and CVE-2026-109619). They used CVE-2026-123039 to get root. In the Israeli attacks they then moved to VMware ESXi hosts and ran a wiper called destroy_v6. Check Point advises installing the November Jumbo Hotfix on all Management Servers and Security Gateways. It also names signs to hunt for: session files containing
upload_progress_pwn, unexplained gaps in firewall logs and SSH from the gateway to internal ESXi hosts. - ScreenConnect CVE-2026-84869. SCKit tried to use this client flaw to spread to technicians. It allows a file to be transferred and run through an active session without the host confirming. It was fixed in 26.6.5 on September 8. Every technician machine should be on 26.6.5 or later.
The forensic read
Two of this week's sources reach for the same word. Barracuda issued its key for a project. Huntress found RMM+ was usually added for a project and left in place. The Qubole roles fit the same pattern: Sentra found them still in place long after customers stopped using Qubole. Infoblox's reported vendor was there to move files, which is a project too. Each time the work ended and the access did not.
That makes the first investigative step the same in every case. Build the life of the credential before building the life of the intrusion. When was it created and for whom? When was it last used for its real purpose? When did it start being used for something else? The gap between the second and third dates is the most useful evidence most of these cases will have. It is also the evidence most likely to be gone. CloudTrail's built-in history keeps 90 days. Microsoft 365's standard audit log keeps 180. A key retired in spring and abused in autumn may have no surviving record of its quiet months unless someone kept the logs on purpose. Keeping them is cheap. It has to be decided before the incident, which is what our Digital Evidence Readiness Self-Audit is built to check.
Sherlock Forensics' view has not changed. An intrusion is not over when the attacker leaves. It is over when someone can show what happened in a form that holds up to scrutiny. If a cloud account, a remote support platform or an AI gateway of yours is in scope for any of these stories, our incident response team can preserve the evidence and build the timeline. An incident response retainer means the first hour is already planned the next time. Next edition, the week of November 21.
Sources
- Barracuda links cloud breach to ARTEX AI-powered cyberattack, BleepingComputer: www.bleepingcomputer.com
- Trusted by default: Qubole cross-account access abused in the wild, Wiz: www.wiz.io
- Inside the Qubole supply chain attack: how Sentra found it, Sentra: sentra.io
- Extortion group claims theft of nearly one billion Expedia traveler records, Security Affairs: securityaffairs.com
- Cyber giant Infoblox confirms security incident after hackers access company information, TechCrunch: techcrunch.com
- SCKit worm installs rogue ScreenConnect clients and session cookie relays across hundreds of endpoints, Huntress: www.huntress.com
- Rogue remote-access campaign spread through ScreenConnect add-on, researchers say, CyberScoop: cyberscoop.com
- ConnectWise ScreenConnect: attackers install malware through RMM+ extension, heise online: www.heise.de
- LiteLLM has a sandbox escape in custom-code guardrail (CVE-2026-40217), GitHub Advisory Database: github.com
- CISA adds four Known Exploited Vulnerabilities to catalog, November 18, CISA: www.cisa.gov
- NVIDIA investigating Gentlemen ransomware breach claims, BleepingComputer: www.bleepingcomputer.com
- Check Point gateways hacked via PHP2Shell to deploy VMware ESXi wiper, BleepingComputer: www.bleepingcomputer.com
Previous in this series: Weekly Security Roundup: November 1 to 14, 2026.
