Patch CVSS 10.0 Cisco SD-WAN Flaw; Mexican Water Utility Attacked Using AI; Another Linux Kernel Privilege Escalation Flaw

These are the droids you're looking for. Don't overlook this as the flaw from March: same daemon, new issue, new patch, no workaround. The Cisco advisory includes the IoCs, a validation checklist, and example output. Check Cisco's fixed releases table, as you may have to move off an EOL version to get the fix. Note that the cloud managed Cisco SD-WAN has already been patched.

I should highlight here that this system in question is the SD-WAN component of the Cisco Product line; if you are a customer, you should patch immediately. There is also guidance on how to see if you have been impacted.

It’s been a bad few months for Cisco. That said, as a founding member of Project Glasswing, hopefully Cisco is leveraging Claude Mythos to accelerate its vulnerability detection and patch management. In the near-term, organizations should look to automate their own patch management process as vendors make patches available.

The cost of patching, not to mention the risk of not doing so, is part of the cost of use. It should be identified (or at least assumed) as part of the acquisition process and decision to buy and use.

The important takeaway here is not that AI suddenly made the attacker unstoppable. The attacker still needed access, persistence, and operational understanding. What AI changed was the speed and scale at which reconnaissance and targeting could happen once inside the environment. There is an important defensive lesson in this story. The attempted password spray against the OT interface failed because basic security controls were working. That reinforces a point many administrators should understand: strong foundational practices still matter, even in an era of AI-assisted attacks.

My buddy John asked me if we're seeing OT attacks on critical infrastructure other than the electrical sector, and here we are. Read the Dragos writeup; their recommendations apply broadly. AI is redefining what we consider "secure enough" — that space where you've implemented enough controls to achieve an acceptable level of risk. There is no such thing as absolute security. Security through obscurity has never been a sound principle either. Tempting, sure, but effective, no. What we need to do is revisit our accepted risks and cyber hygiene with an eye to AI, keeping in mind that basics like patches, limiting access, MFA, password hygiene, and comprehensive hardware (and software) inventory still work.

This story is a timely reality check on AI-assisted cyberattacks. While the headlines and hype will show that AI appears to have accelerated reconnaissance and attack preparation, traditional security controls such as strong passwords still proved effective in preventing further compromise. Given the rise in this type of attack, organisations should avoid both being overly complacent and being distracted by the hype. Remember that AI may improve attacker efficiency and scalability, but good foundational security practices remain highly effective.

Imagine how long a manual attack like that might have taken. AI is a productivity enhancer. That said, this attack required significant special knowledge on the part of the attacker. A human being is responsible for anything the computer is asked to do and for all of the results.

Expect more AI-assisted attacks in our future. The best defense remains adherence to strong cybersecurity fundamentals such as the CIS Critical Security Controls. If you make that investment, you’re better positioned to absorb the change brought on by AI. If you haven’t, no time like the present to begin.

Fortunately, the mitigation for Dirty Frag, disabling esp4, esp6 and rxpc modules, also works for Fragnesia. This flaw modifies cached file pages, not on-disk files, using created namespaces with CAP_NET_ADMIN privileges. In this case you need to watch for suspicious namespace creation, XFRM manipulation of use of AF_ALG. If detected, clearing page cache contents or rebooting will clear out the modified in-memory binaries.

There will be many more of these coming I believe. Here is the thing to note about this one: there is currently no patch but there is a workaround. If you can disable this feature in the kernel, then maybe you should try and do that. I know the kernel team is working to try and get a kill switch for modules and functions as a workaround for these types of scenarios.

For whatever reason, the vulnerability researcher(s) don’t feel compelled to follow the responsible disclosure process. That’s unfortunate, as the choices available to you until the major Linux distributions offer a patch are somewhat stark.

As expected, using AI for discovery and resolution of vulnerabilities is resulting in a spike in flaw counts. Microsoft, Apple, Mozilla, Google and Oracle are using AI, likely as part of Project Glasswing, for flaw discovery and remediation. That is a lot of flaws to digest/understand. Ultimately, the number of newly discovered flaws is expected to taper off; fortunately, resources such as Rapid7 and the SANS ISC are helping with the heavy lifting when it comes to finding the more concerning issues. Specific to Microsoft, you want to keep an eye on CVE-2026-41089, a Netlogin flaw which allows SYSTEM privileges; CVE-2026-41096, a Windows DNS RCE flaw; and CVE-2026-41103, a privilege escalation flaw using forged credentials bypassing Entra ID.

About half of the critical flaws only affect Microsoft’s cloud products, and users do not have to take any action. Microsoft, in recent months, has become more transparent about vulnerabilities in its cloud products, leading to more of them being included in “Patch Tuesday.”

Validates my long-held contention that Microsoft legacy code contains a reservoir of both known and unknown vulnerabilities. For reasons that include backwards compatibility, they are difficult to fix. That said, moving them from unknown to known is a necessary first step.

While attacks disrupting the supply chain are nothing new, we are seeing more and more of them occurring. In today's world of complex supply chains and just-in-time manufacturing, businesses need to ensure that their resilience planning looks at not just the cybersecurity risks within their supply chain but also at the disruption a cyberattack on a key supplier could have on their business, and implement controls to mitigate that impact.

Foxconn is targeted as a manufacturer and an opportunity for supply chain disruption or manipulation/theft of IP. The Nitrogen Ransomware gang is taking credit for the attack, claiming to have taken 8TB of data including technical drawings, instructions and internal product documentation for products from Apple, Dell, Intel, Google, Nvidia and others. While the specific attack vector is unknown, Nitrogen specializes in SEO poisoning and fake software downloads to distribute malware. Make sure that your EDR is ubiquitous, configured for maximum effectiveness, and that you are actively blocking malicious sites. Ensure that downloads are verified, and that you can detect and stop attempted exfiltration of data.

Not a lot of details, but it just goes to show that even a Fortune 30 company can fall victim to a suspected ransomware attack. It’s unlikely that we’ll know the details of the attack, and more importantly, what cybersecurity defenses were lacking.

This incident demonstrates the growing risks associated with the unsanctioned use of AI tools within organisations. Just like the introduction of BYOD through smartphones, people are going to use these AI services unless structured and workable enterprise AI alternatives are made available. Banning the use of AI tools will not work, however companies need to understand that sensitive data being entered into unapproved AI services can quickly create both security and regulatory issues. Organisations, especially those operating within the EU and regulated by the EU GDPR and the EU AI Act, should ensure clear policies exist around the use of AI applications, particularly where personal or sensitive data is involved. Introducing authorised and sanctioned tools that staff can easily use is also a necessary step. Blocking access to public AI services via web proxy solutions and/or monitoring DNS traffic going to such sites could be an additional layer of control.

With the intense pressure to adopt and embrace AI, guidance is essential. It is really tempting for staff to use their credentials to point an AI agent at data to conduct analysis without consideration of where that data is copied/processed. Our challenge is to help staff use AI securely, to support delivery and innovation without compromising security or causing undue delays in delivery. Take a hard look at creating local instances, to include your private cloud services, where you can control data flow and access. Do not become the problem (or roadblock) that the smart people you've hired feel the need to solve.

Here’s an example of an employee’s desire to use AI having harmful effects on the company. An AI Acceptable Use Policy coupled with employee AI awareness training could have reduced the likelihood of this data spillage. Let this serve as a reminder to companies to create policies around the use of AI for business operations.

As organisations increasingly deploy AI systems, understanding the provenance of models, datasets, and supporting infrastructure becomes essential for both security and governance. Organisations operating in the EU should pay close attention to these developments given the wider regulatory focus under laws such as the EU AI Act and the EU GDPR.

We're all trying to figure out what questions to ask about AI to understand and secure it. This document is only 26 pages long and is worth reviewing. While we don't yet have AIBOMs, this is a minimum set of characteristics you'll want to understand about AI systems, which you can then leverage to ensure you've appropriately incorporated necessary cybersecurity tools. Keep an eye on this as it matures.

The re-use of poorly documented code is risky. Developers are responsible for all the code they use in their product, whether or not it is original. A digital SBOM is essential documentation.