SEC536: Adversarial AI - Penetration Testing AI Systems


Experience SANS training through course previews.
Learn MoreLet us help.
Contact usBecome a member for instant access to our free resources.
Sign UpWe're here to help.
Contact UsSomething mysterious is rising from the ice near Santa's castle, and the elves need YOUR help! The 2026 SANS Holiday Hack Challenge: The Pyramid on the Ice launched Wednesday, November 18, inviting cybersecurity enthusiasts everywhere to explore a whimsical winter wonderland filled with fun mysteries and hands-on cybersecurity challenges. Players of all skill levels can sharpen their skills in AI security, web security, cloud computing, reverse engineering, and more. There's even a new award for the most creative use of AI! Best of all, it's FREE, SANS's annual holiday gift to the cybersecurity community, 23 years running. Learn new skills, play with friends, and compete for prizes! The challenge runs through January 5, 2027. So grab your virtual mittens, fire up your terminal, and help the elves!
Read more and play at Holiday Hack Challenge 2026: The Pyramid on the Ice
—Ed Skoudis
SANS Chief Holiday Officer
On November 17, researchers at Calif released their report on a high-severity flaw affecting open-source implementations of the Java programming language, and Oracle released a fix on the same day. Dubbed "Java4Shell," CVE-2026-114938 carries a severity score of 8.1 in CVSS 3.1, and 9.2 in CVSS 4.0. The vulnerability allows an unauthenticated attacker to crash the Java Virtual Machine and/or achieve remote code execution by exploiting a bug in the Base64 decoder of the Open Java Development Kit (OpenJDK) 17.0.6 and newer to cause a heap-based buffer overflow and out-of-bounds write. Calif "discovered a bug in [x86-64 CPU] optimized machine-level code that lets malformed input overwrite neighboring memory ... with careful exploitation," including trigger input as simple as a string of the character "A" 64 times followed by an equals sign. Calif's writeup also states that "there is no universal payload. Each chain needs a decoder route, a live writer, a returned reader or another feedback path, a heap layout, and an application action. Every target needs a bespoke profile. The exploit is not a universal point-and-shoot." The US Cybersecurity and Infrastructure Security Agency (CISA) has urged organizations to inventory affected Java installations "including those bundled with third-party applications or included in container images and appliances," apply patches from Oracle and from vendors, review Oracle's security advisory, and monitor for rogue runtimes, unexpected JVM crashes, and suspicious Java processes.

It’s important to note that the exploit is application specific, meaning this is going to take a bit of work to exploit. As Johannes said, this isn’t an INFOCON yellow or red alert — merely condition green, meaning the internet is in its usual “broken" state. Work your normal 9-5 day, and what you need to do is make sure you’ve queued up the Java updates, particularly to your low-risk, auto-update patching tiers. Earlier versions of this alert indicated updates were not available yet; the November 20 CPU from Oracle includes the updated JVMs. A mitigation is to disable AVX-512 on 64 bit JVMs on X86-64 hosts _only_; as this is one of those where the mitigation works in certain configurations, I’d focus on updating the JVM instead. Then have your threat hunters check for the IoCs in the SentinelOne blog. You may wish to implement JVM behavioral monitoring to detect unusual activities conducted by your JVM. Don’t blow the weekend working this one.

Another highly complex bug to exploit. It requires specific hardware, ideal network conditions, the exploit has to be tailored to the application, and on and on. We won't see mass exploitation across broad environments on this bug, but we might see mass exploitation on one specific type of device, like we saw with PHP2Shell (where mostly network firewalls were targeted). It's the latest example in what is becoming a regular series of "celebrity" vulnerabilities that are more valuable for media clicks (this bug in particular had a coordinated NYTimes article) than they are for attackers. These bugs are increasingly wasting defenders' time, and if they continue at this rate, asset owners will have to start deferring them. We can't be initiating rapid response 3+ times every week without burning out both tech teams and creating more large-scale outages. We have to remember — the goal isn't to make the business as secure as possible, it's to protect the business from disruption, risk, and losses. When the patch is more disruptive than a potential attack, WE become the problem.
Calif
SentinelOne
SecurityWeek
The Register
SANS ISC
Help Net Security
The Hacker News
CISA
Earlier this week, websites at least 178 organizations were found to be serving malware that skimmed payment card data, exfiltrated cookies and screenshots, swapped out cryptocurrency wallets for one designated by attackers, and delivered iOS exploit kits to users viewing the sites on iPhones and iPads. According to Optimizely's initial report, "on 16 November 2026, an unauthorized third party exploited a vulnerability in the Optimizely Web Experimentation account signup flow. They used it to impersonate users and insert malicious JavaScript into the Project JavaScript of some customer projects. This was not the result of a compromised Optimizely administrator account." Affected organizations and websites include: Air France-KLM, the BBC, CNN, Equifax, GitLab, HSBC, IKEA, HSBC, Ledger, Peacock, Salesforce, and Zoom. A write-up from Report URI notes that of affected sites, 68 had no Content Security Policy (CSP); 50 had CSPs but still allowed the payload to deploy and the data to be stolen; 17 had CSPs that allowed the payload to run but blocked data theft while still allowing crypto wallet address swapping; and 24 had CSPs that blocked the payload. Researcher Scott Helme notes, "The important lesson here is that trusted JavaScript doesn't always remain trustworthy. A website can be compromised through a third-party dependency without its own code or infrastructure ever being breached." The Report URI write-up includes IoCs as well as advice for sites that use Optimizely and for users who visited affected sites.

I know some of you are wondering if you can disable JavaScript on your site and your browser. Typically sites have 20 or so little scripts which profile the user and browser to provide a targeted experience as well as capture user data, for whatever purposes. The reality is, most users aren't up for a noscript or otherwise disabled JavaScript experience; it's their ad blocker which is at play here, coupled with your site's content security policy. And with ads or analytics, you have elements outside your control being delivered. Focus on what you can control. Make sure that the CSP is not allowing scripts to fire before consent is granted and that any JavaScript is the approved one. There are a number of headers and other security settings, in addition to the CSP, that we need to ensure are in place; make sure someone has the task of making sure you’re staying on top of these.

One way to secure your supply chain is to reduce your exposure. Is Optimizely worth it? In particular, considering the risk of exposing your customers to malicious code? Content Security Policies may help, but you may have opened them up to allow for tools like Optimizely to function properly.

Consumers' risks from online commerce are going up. They should be encouraged to use payment proxies (e.g., PayPal, Apple Pay, Google Pay) and one-time and one-merchant tokens (e.g., Privacy.com).
Krebs on Security
Ars Technica
The Register
Report URI
Optimizely
On November 18, 2026, F5 published ten security advisories addressing vulnerabilities in open-source web server software NGINX, five of which have been dubbed "NGINX2Root" by the researchers at Xint who discovered them, specifically when any one of these five is independently chained with a separate vulnerability found in the Linux kernel. All five of these flaws carry a CVSS 4.0 score of 9.2. CVE-2026-123007 is considered the "most impactful," as it allows an unauthenticated attacker to crash the NGINX worker process, corrupt its memory, or remotely execute code "on systems with Address Space Layout Randomization (ASLR) disabled or when the attacker can bypass ASLR" by causing a heap buffer overflow and an NGINX worker segmentation fault on NGINX Plus or NGINX Open Source configured to use HTTP/2. Xint was able to chain this flaw "with a separate flaw in the Linux kernel's SCTP subsystem to gain root privileges on the host." The other four vulnerabilities are all use-after-free flaws. CVE-2026-123634 is an issue in NGINX's DNS resolver logic allowing the same malicious results as CVE-2026-123007. CVE-2026-123633 and CVE-2026-121810 both allow denial-of-service or code execution; the former is "caused by an active HTTP/2 connection remaining eligible for draining," and the latter is "caused by closing a QUIC connection while one of its HTTP/3 requests is still running." The last of the five, CVE-2026-113425, allows remote code execution through a flaw in the mail proxy's auth_http authentication path. All NGINX releases from 1.9.5 to 1.31.8 are affected, and users are urged to update to NGINX version 1.31.9 for the complete fix; Xint also notes that CVE-2026-123007 and CVE-2026-123633 can be blocked by disabling HTTP/2, and offers emergency mitigations to apply for the other three flaws until the patch can be applied.

None of the vulnerabilities are easily exploited. A theme emerging in web servers is memory-handling flaws in HTTP/2. One workaround is to disable HTTP/2, which may be advisable, given that researchers are pointing their AI tools at this code and they may discover more flaws soon. However, disabling HTTP/2 carries its own risks, particularly if you have middleboxes that terminate HTTP/2 and forward HTTP/1.1. HTTP request smuggling is a common vulnerability in this scenario.

Hopefully your IT team has already reached out to any of your hosting services to get their status on this update. A year ago we could probably disable HTTP/2 without a lot of concerns, but these days, it’s likely to backfire. Don’t overlook appliances which have NGINX running, like your WAF, load-balancers, and even network gear. Yeah, I know, it's like everything has a web server these days, which is why we want to protect those management interfaces. The good news is that there is no active exploitation, but with all the press, that’s likely to change, so scan your inventory, identify your NGINX targets, and get your team rolling the updates — prioritize internet facing services.

This one serves as a reminder that vulnerabilities in high-market-share open-source software (think WordPress, Apache, Drupal, Joomla, etc.) are high leverage attack points against you and your supply chain. The wave of AI-discovered vulnerabilities is more like a rising tide rather than a one-time tsunami — more/faster updating is going to be needed. That relies on accurate and fresh software asset/configuration management.
F5
GitHub
Xint
The Register
CyberScoop
BleepingComputer
SecurityWeek
The Hacker News
SANS ISC
On Thursday, November 19, 2026, the PHP foundation released PHP 8.6 and updates for all supported versions of PHP (8.5.13, 8.4.28, 8.3.37 and 8.2.36), including security fixes for critical flaws that only affect "either non-default configuration or uncommon setups." The updates address 18 flaws in total, and the seven critical flaws all carry CVSS score 9.2. Aikido security highlights CVE-2026-129289, which allows an unauthenticated attacker to bypass front-end routing and execute PHP code with worker privileges by forming a FastCGI request with unread bytes from an HTTP request body that PHP failed to discard before reusing a persistent PHP-FPM or LSAPI connection. Exploitation of this flaw requires "a front end that enables persistent backend connections and an application path that opens php://input without reading it fully. Relevant configurations include nginx with fastcgi_keep_conn on, Apache mod_proxy_fcgi with enablereuse=on, HAProxy FastCGI applications that retain backend connections, and LSAPI." Configurations that close the FastCGI connection after each request are not affected. CVE-2026-115560, a session serializer mismatch that can lead to remote code execution, was initially disclosed by PHP maintainers on October 23 when they recommended that users "set session.upload_progress.enabled=0 until the next regular bugfix release" to cover a possible vector for RCE. The other five critical flaws addressed are three use-after-free vulnerabilities allow remote code execution after interrupted session shutdown (CVE-2026-129481), through repopulation of the last fatal error backtrace global after shutdown (CVE-2026-129182), and in in magic method backtraces (CVE-2026-129474); a time-of-check time-of-use (TOCTOU) vulnerability that uses partially overwritten session files to allow remote code execution (CVE-2026-126714); and finally another use-after-free in the timezone cache that allows remote code execution (CVE-2026-129496).

18 PHP flaws addressed, some of which can be used for RCE in nonstandard configurations. This looks to be a use-after-free flaw, and you can focus on those — but you still have PHP with the vulnerable code should the configuration change — or just get everything updated, which is a lot easier to explain to non-technical folks. Look at it this way: We just went through PHP2Shell fixing all our PHP, so we know where it is, and who has to address things. And yes, you’re absolutely right in feeling you just finished painting this bridge; if it makes you feel any better, I'm making sure my coffee supply is sufficient.

More cleanup from PHP. The vulnerability that caught my attention isn't the highlighted remote code execution vulnerability, but another one that switches from HTTP to HTTPS if “HTTPS” is capitalized in the URL. It made me consider if this issue is present in other languages/libraries as well.
PHP
Aikido
BleepingComputer
SecurityWeek
The Hacker News
While last week's M365 outage and the FAA ground stop in the US may raise concerns that these outages are more than a coincidence, many experts point to other explanations: Cynthia Kaiser, senior VP of Halcyon’s Ransomware Research Center and former deputy director of the FBI’s Cyber Division, said the incidents demonstrate "how deeply interconnected our digital infrastructure has become, and how a single technical failure or software update can cascade with outsized effects across critical systems." Mike Chapple, University of Notre Dame Professor of information technology, analytics, and operations said, "When we dig into the details, each one has an explanation that makes sense on its own, but they all point to a common theme: the complexity and fragility of the systems that we depend on every day." NSANet, the US National Security Agency's (NSA's) classified internal communications network, experienced disruptions earlier this week, prompting the agency to implement delayed opening and stagger employees' arrival times. In a statement, an NSA spokesperson said that NSANet "experienced temporary degradation during scheduled routine maintenance. While some internal systems were affected, mission-critical capabilities remain fully operational." The outage appears to have been due to issues with PHP-related updates. Earlier this week, Apple experienced some system outages following an internal update, causing issues with ApplePay and iCloud, and JP Morgan Chase acknowledged that it "experienced a temporary issue related to a software bug in on-premise hardware systems;" some customers reported being unable to use their ATM cards and or conduct transactions at brick-and-mortar bank locations. Several other financial institutions and United Airlines also reported short-lived disruptions.

Over the last year, when talking to students in class, the trend is clear: We not only overcomplicated systems by relying on more and more external resources and complex supply chains, but at the same time, executives are desperate to show the promised ROI by cutting staff, which they would actually need more of to manage the complexities they introduced. The only thing that surprises me each morning is that the lights still come on.

The days of running systems for long intervals without service interruptions for security updates are long past, although I remember when we could run a system for a year without memory leaks or other stability issues. Also past are the times where we could spend months regression testing before impacting production. Instead, we need to not only have our prioritized/tiered update plans, but also a clear understanding of whether our services fail over or connections are migrated without impact during maintenance/downtime, and then implement maintenance accordingly. Have a tabletop of how you can not only take components offline without unacceptable impact but also ensure the updated components are suitable for production workload before bringing them back online. It may be you need a bit more testing before declaring the updated system is ready. The good news is there are automated harnesses which make this much better than calling a staff person to start clicking on the UI.

This one is good incentive for an extra credit exercise: If you have backup plans for outages in any of your major service providers, do you know their common dependencies (like cloud service providers), and have you game-planned for a scenario where two of them go out at the same time?
These isolated events have sparked plenty of theories, from nation-state attacks and government cover-ups to simple tech glitches. The real culprit? AI. Using AI to find software flaws faster is forcing teams to rush out and apply patches, and speed always brings risk. Take a deep breath, we'll get through this together.

The first rule of change management is, "If it worked yesterday, it must work today." Changes should be made in such a way that the system can be returned to a known working state in minutes.
WIRED
The Register
Gov Infosecurity
Nextgov/FCW
The Record
NYTimes
NYTimes
The People of the State of California have brought a lawsuit against OpenAI on the grounds that the company knowingly developed AI models without proper safeguards and monitoring in place, and thereby misled and endangered the public. The suit opens by citing the attack on Hugging Face that took place in July 2026, alleging that it was "the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability," and arguing that the company withheld information and obstructed investigation by the Office of the Attorney General after admitting to more than 80 incidents in which it lost control of autonomous AI agents. Ars Technica notes that OpenAI had only disclosed 27 of those incidents publicly. California Attorney General Rob Bonta contrasts "science fiction fantasies of 'rogue' AI systems run amok" with his contention that "the attack was entirely foreseeable in light of OpenAI’s decision to test models it knew to be dangerous in a flawed and exploitable testing environment without proper oversight." The plaintiffs claim that OpenAI's conduct violated California's Unfair Competition Law with "unlawful, unfair, or fraudulent acts or practices" through misleading statements, through the company's role in agents gaining unauthorized system access through fraudulent means, and through unfair acts comprising negligence around monitoring and safeguards. The plaintiffs also claim OpenAI broke the law by creating a public nuisance that now must be abated, alleging that its practices "harmed, and continue to endanger, public property, safety, peace, comfort, and convenience in California." The suit seeks to enter an injunction that prevents the unfair business practices, to see restitution paid for money or property acquired by means of those practices, to award civil penalties, to award remedy of disgorgement, to see legal fees repaid, and to compel OpenAI to abate the public nuisance.

I’m not a lawyer, nor do I play one on TV, but this concerns me. I stand by my statement that you need to be accountable for your systems' actions, AI or otherwise. It's not clear that these actions will benefit the injured parties, but they could result in added guidelines and best practices to prevent future incidents. The trick is how those will map to independent researchers using open source AI models and how they will interplay with the existing federal guidelines and frameworks for AI designed to ensure responsible development and deployment.

The state of Florida also sued OpenAI earlier this year. Regardless of the outcome of these (and future) lawsuits, the legal fees alone will likely be greater than Open AI’s cost of making sure the test environment was really, really, really isolated and safe for the test.
Startups are famous for moving fast and cutting corners to get new features out the door. The game-changer this time? Two words: artificial intelligence. That's why California filed its novel lawsuit, and why other states will likely copy it. Fast-forward through years of legal pushback, court delays, and jury selection, and both sides will finally settle around 2030. By then, the AI models at the center of it all will be unrecognizable.

In civil litigation, third parties may have the burden of proving standing and damages. We need Hugging Face to join CA and FL in these suits. Their standing and damages are obvious. If we are to enjoy the benefits of AI while avoiding unintended consequences, it is essential that we hold its users responsible for their use. OpenAI was clearly reckless.
California OAG
California OAG
Ars Technica
The Guardian
The Record
BleepingComputer
The European Commission is recovering from a cyberattack that began last month and targeted the organization's internal IT systems. The incident was disclosed in a press briefing on Monday, November 19, in which Commission spokesperson Thomas Regnier said suspicious activity was detected on October 22 and confirmed to be a cyberattack on October 26. The attack has affected the European Commission's translation systems, as well as its Sysper system, which staff members use to access pay slips and handle timesheets, leave requests, and medical reimbursements. Politico, which first reported the incident, said that CERT-EU indicated the attackers used the ARTEX AI-powered penetration testing tool. The US Federal Bureau of Investigation (FBI) issued a Flash alert last Friday warning that ARTEX was being used maliciously. The European Commission has experienced two other cyber incidents this calendar year: A January attack on its mobile management infrastructure was linked to vulnerabilities in Ivanti Endpoint Manager Mobile, and in March, the Commission's Amazon Web Services (AWS) account was breached and data were stolen.

Mobile device management infrastructure is a target. Ivanti EPMM or otherwise, make sure not only that you're keeping it updated, but also that you have deployed the most current security guidance, to include restrictions to management interfaces and APIs as well as MFA. Don't fixate on the tool, such as ARTEX, being used to discover and exploit vulnerabilities, focus instead on timely cyber hygiene. The fixes aren't changed, the risk profile has, to include an increased likelihood of success. You’ve got that list of round-to-it items — it's time.
The Register
The Record
Infosecurity Magazine
The Record
The Record
On Tuesday, November 17, Norwegian multinational energy company Equinor disclosed that it detected a cyberattack against its IT systems. The company says that the attack has not affected the Equinor's production, and it has reported the incident to police and is working with national security authorities. While Equinor has not provided specific details of the attack, it noted that it "is not a virus attack, [so] there is no risk of the attack spreading." As of an update on Thursday, November 19, Equinor said the attack was "ongoing." Reuters learned from an anonymous Equinor source that the company's internal email system was down.

Read the Equinor update on the attack. They are responding rapidly, restoring systems while still containing and thwarting attackers. Not only that, they detected and responded to the attack in a timely fashion. While more details on the attack would be helpful, my guess is they are being hit with ongoing DDoS activities targeted at their IP ranges rather than at them specifically.
Equinor
CBC
The Register
The Record
Reuters
BT
Transport for London (TfL) is reporting that a breach of its Oyster photocard website has exposed personal information of roughly 265,000 Oyster photocard holders. TfL operates much of London's public transportation, including the Underground, buses, and trams. TfL learned of the incident on November 16 after customers reported receiving email asking them to create accounts. The Oyster photocard website is temporarily offline while the incident is being investigated. Compromised data include names, phone numbers, email and street addresses, dates of birth, and security questions and answers. TfL has contained the incident, notified affected individuals, and notified the UK Information Commissioner's Office (ICO). A breach at software provider Wellington IT has compromised sensitive data, including payee names, account numbers, and addresses, at 40 Irish credit unions. Robert Troy, Ireland's Minister of State at the Department of Finance with responsibility for credit unions, said that while the incident occurred in October, Wellington IT is only now notifying affected credit unions. In a statement to news outlets, Wellington IT said that it notified customers "as soon as [it] confirmed the potential risk to customer data."

Smaller financial institutions, particularly credit unions, use service providers, often called Credit Union Service Organizations (CUSOs) for services, which provides a huge advantage over creating, managing, and securing the service themselves, but it also opens them up to the risk of third-party compromise. They are faced with the same challenges the rest of us have, not only of verifying the security of those CUSOs, but also of ensuring that incident notification and response meets their requirements. Believe you me, their regulators make sure these are not skipped or overlooked. One thing I've started to see is credit unions and financial institutions offering ID/credit monitoring as part of some types of accounts, or as a separate service, usually at a much lower cost than we can get on our own. This means that while we, as members of a Credit Union, or any other financial institution, can't do anything about their accepted third-party risks, we can leverage these services to get a needed service at a reduced rate to protect ourselves.
Most of us are numb to breached data and free credit-monitoring offers at this point. Organizations still hold back on explaining how the attack happened and which security weaknesses were exploited. Better information sharing turns painful breaches into lessons everyone else can use.
The Register
BleepingComputer
The Register
BleepingComputer
On Monday November 16, the US Department of Health and Human Services (HHS) issued a final rule that modifies the Standards for Privacy of Individually Identifiable Health Information under the Health Insurance Portability and Accountability Act of 1996 (HIPAA) and the Health Information Technology for Economic and Clinical Health (HITECH) Act. The modifications focus on three aspects of care: "strengthening the Right of Access to PHI, facilitating care coordination, and reducing the regulatory burden on covered entities." Covered entities will now have 15 days to provide individuals with copies of their PHI. The rule also allows individuals to direct covered entities to provide copies of their ePHI with apps designated by the individual, provided the covered entity is capable of providing the information in the required format. The rule also grants covered entities permission "to disclose PHI for individual-level care coordination and case management to certain third parties." Prior to the final rule's publication, a letter from the College of Healthcare Information Management Executives (CHIME) and a coalition of healthcare organizations to HHS expressed "concerns about the ability of providers to confirm that a patient authorized a third-party app to receive their PHI." The new rule will take effect on Monday, January 18, 2027.

This was expected back in August, albeit the first publication of the NPRM proposed changes to the HIPAA privacy rule was six years ago — it's now here, published in the Federal Register today, with an implementation timeline of 60 days. Translation: You must implement by January 18. Mind you, the updated rule is now 363 pages, so you're going to need to read selectively. There are changes which may impact you, such as refinements to the definition of an electronic health record, or a shortened timeframe for providing individuals with copies of their PHI (now 15 days versus 30) delivered to the health app of their choosing, provided it has that capability. Let others worry about what didn't make it into the final ruling; with that timeline, you're going to work with your service providers on meeting the requirements which are present. Leverage HHS's fact sheet to focus your research.

Seeing increased access by healthcare apps paired with vague terminology about not allowing "unreasonable" identity verification worries me — the CHIME letter to HHS captures the issues nicely. Example: Making sure requiring the individual authorizes access is key to blocking rogue apps.
The updated rules modernize data sharing and make personal health information (PHI) easier to transfer, but they come with a security risk. The security tradeoff is in the use of personal health applications to retrieve the PHI. Adversaries will look to exploit those sorts of applications, likely taking advantage of poor consumer patching practices. That's difficult for HIPAA rules to protect against.

These rules make legal what five of my providers already do anyway.
HIPAA Journal
Gov Infosecurity
Federal Register
HIPAA Journal
SANS Internet Storm Center StormCast Friday, November 20, 2026
Honeypot Analysis as Code; PHP Patches and Session RCE (CVE-2026-126714); Dragos AI Canary Tokens; vCenter Exploit (CVE-2026-123890) (Nov 20th 2026)
https://isc.sans.edu/podcastdetail/10188
This episode covers a SANS.edu intern's Git-and-pull-request workflow for AI-assisted honeypot analysis, today's PHP security updates including a session-handling remote code execution (RCE) bug, Dragos research that planted canary tokens in AI prompts, and an already exploited VMware vCenter vulnerability.
Treat Your Honeypot Analysis Like Code: Git, Pull Requests and a Human Gate for AI [Guest Diary]
SANS.edu undergraduate intern Kevin Truong explains how he treats honeypot analysis like code, using Git and pull requests so an AI assistant can propose query and report changes while a human analyst reviews and approves them. He says the same workflow can apply to other kinds of log analysis, not just honeypots.
https://isc.sans.edu/diary/33534
PHP Vulnerabilities and Patches
PHP released security updates in versions 8.5.13, 8.4.28, 8.3.37, and 8.2.36 that fix memory-corruption issues leading to remote code execution (RCE), including a session-handling flaw, CVE-2026-126714, that depthfirst documents as phpAfterShell. A separate moderate issue, CVE-2026-128064, can cause fopen() and similar stream functions to fetch an HTTPS or FTPS URL over plaintext when the scheme uses an uppercase S.
https://www.php.net/archive/2026.php
https://github.com/php/php-src/security
https://depthfirst.com/research/phpaftershell
Supply Chain Canaries
HD Moore, now with Dragos, describes planting canary tokens, also called honey tokens, in prompts sent to AI services in order to see whether those secrets later leak. Dragos sent approximately two million such requests and found that at least three AI providers exposed the tokens after apparent internal compromises, that inference-marketplace proxies inspected and used some of the tokens, and that some organizations did not honor their privacy policies.
https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
VMware vCenter Exploit
Mandiant and Google Threat Intelligence report that CVE-2026-123890, an unauthenticated authentication bypass in VMware vCenter Server, is already being exploited. The bug is in the vSphere stack's Directory Service and can yield an administrator session; Broadcom released patches in advisory VMSA-2026-0009 and lists no workaround.
https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere
No Podcast Thursday and Friday
Next week is the Thanksgiving holiday, so Stormcast will only be published Monday through Wednesday.
SANS Internet Storm Center StormCast Thursday, November 19, 2026
NGINX2ROOT HTTP/2 RCEs (CVE-2026-123007, CVE-2026-123633); Microsoft Agent Framework PyPI Compromise; Holiday Hack Challenge
https://isc.sans.edu/podcastdetail/10186
Today's episode covers F5's NGINX2ROOT HTTP/2 remote code execution patches, a malicious release of Microsoft's Agent Framework on the Python Package Index (PyPI), and the launch of the SANS Holiday Hack Challenge.
NGINX2ROOT HTTP/2 Remote Code Execution (CVE-2026-123007, CVE-2026-123633)
F5 patched five NGINX vulnerabilities, including critical HTTP/2 memory-corruption bugs that can allow unauthenticated remote code execution. The most widely exposed issue, CVE-2026-123007, does not require a special NGINX configuration beyond HTTP/2 being enabled, though exploitation needs a large volume of traffic and is not always reliable, even when Address Space Layout Randomization (ASLR) can be bypassed. Researchers showed these bugs can be chained with a kernel privilege-escalation vulnerability to obtain root access. Rapid7 released a proof-of-concept that works against NGINX running in containers; until you can patch, consider turning off HTTP/2 (Hypertext Transfer Protocol version 2).
https://my.f5.com/manage/s/article/K000163926
https://www.wiz.io/blog/nginx-http2-rce-vulnerabilities
https://github.com/sfewer-r7/CVE-2026-123007
Microsoft Agent Framework Compromised on PyPI
Wiz reported that malicious versions 1.26.1 and 1.26.2 of Microsoft's Agent Framework were published to the Python Package Index (PyPI) on November 17 after a compromised publishing credential was used from a Microsoft IP address. Both malicious versions have been removed; 1.26.0 is the latest safe release. The package normally sees about 41,000 downloads a week, and the malicious builds were only available for part of that day.
https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi
SANS Holiday Hack Challenge
The SANS Holiday Hack Challenge is live again, with links in the show notes and across the SANS website. It may be a welcome distraction between patching and incident work.
https://www.sans.org/cyber-ranges/holiday-hack-challenge
SANS Internet Storm Center StormCast Wednesday, November 18, 2026
Phishing Kit Collects MX Records; Java4Shell OpenJDK RCE (CVE-2026-114938); Optimizely Magecart Compromise (Nov 18 2026)
https://isc.sans.edu/podcastdetail/10184
In today's StormCast, Jan Kopriva describes a phishing kit that collects MX records, Calif discloses Java4Shell remote code execution in OpenJDK's Base64 decoder, and an Optimizely compromise spread Magecart-style malware to major websites.
A Phishing Page Interested in MX Records
Jan Kopriva of the SANS Internet Storm Center describes a phishing page that customizes itself based on the victim's email domain, then uses JavaScript to perform a DNS-over-HTTPS (DoH) lookup of the domain's Mail Exchange (MX) records via Google DNS. The MX records are stored with the stolen username and password rather than used to customize the page.
https://isc.sans.edu/diary/33526/
Java4Shell: Write Once, RCE Everywhere
Calif, the firm behind PHP2Shell, disclosed Java4Shell (CVE-2026-114938), a memory-corruption bug in OpenJDK's Base64 decoder on x86-64 CPUs that support the AVX-512 VBMI extension; ARM processors are not affected. The flaw can allow unauthenticated remote code execution in applications such as Keycloak, and can also leak memory in a manner similar to Heartbleed; Oracle released a patch in its November Critical Security Patch Update.
https://www.oracle.com/security-alerts/cspunov2026.html
Optimizely Compromise
Attackers compromised Optimizely, a JavaScript experimentation platform used by many large websites, and swapped the distributed script for a Magecart-style payload that stole payment card and cryptocurrency wallet information. Report URI found the malware on sites belonging to 178 organizations, including HSBC and the BBC; Johannes suggests blocking optimizely.com and the .icu top-level domain.
https://blog.report-uri.com/worlds-largest-brands-infected-with-malware-via-optimizely/
My Upcoming Classes
Catch up on recent editions of NewsBites or browse our full archive of expert-curated cybersecurity news.
Webinar | SANS 2026 Detection and Response Survey Report: Signals, Not Noise | Wednesday, December 9 | Josh Lemon | Explore a data-driven look at how security teams are evolving their detection, triage, and response strategies in an environment shaped by AI-driven threats, ransomware, supply chain attacks, and cloud complexity.
Webinar | The Unpatched Long Tail: What Your Patch Program Misses | Thursday, December 3 | Explore how to uncover hidden software exposure, understand the risk it creates, and decide what to patch, update, replace, or remove.
Webinar | Microsegmentation Buyer's Guide: How to Evaluate Solutions for Zero Trust and Hybrid Cloud | Friday, December 4 | Learn about workload discovery, application dependency mapping, policy design, and enforcement across hybrid environments.
Webinar | The Agentic SOC: Machine-Speed Detection, Investigation, and Response with AI Agents | Monday, December 7 | Dave Shackleford & John Murray