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 UsMicrosoft now says that a bug introduced in a July 2026 networking gateway servers update was responsible for the Microsoft 365 outage late last week that caused more than 80 percent of its network servers to become unresponsive. The incident, which began on Thursday, November 12 and lasted nearly 24 hours, affected multiple Microsoft 365 services, including Outlook, Exchange Online, Teams, and Copilot. BleepingComputer writes that in a preliminary post-incident report in the Microsoft 365 Admin Center (currently available via Reddit), Microsoft attributes the outage to three factors: "a previously undetected software bug, increased traffic after a Microsoft Teams failover, and reduced server capacity." In the report, Microsoft acknowledges that its "normal load testing did not detect the latent issue." Microsoft also notes that about 14 percent of the gateway's server capacity was already offline due to a separate issue. WIRED notes that the incident "caused widespread disruption to schools, courts, governments, and law enforcement agencies—plus countless businesses and millions of employees—around the globe, highlighting how many of the world’s most crucial systems rely upon technologies from a small number of tech companies." Impacts of the outage were felt around the world in multiple sectors: some UK police forces were unable to accept online crime reports; Germany's Deutsche Bahn reported difficulties sending timetables; and other countries saw reports of communications services issues from governments, courts, and healthcare organizations. Health-ISAC CSO Errol Weiss told Gov Infosecurity that members of that organization communicated with staff through Everbridge, AlertMedia, WhatsApp, Signal and text messages.

While this may feel like a captain obvious comment, between MS and the FAA we really need to not underestimate the impact of network changes. The side effects — particularly with so much outsourcing — of a small outage can still be devastating. We really cannot put the cloud/outsourcing genie back in the bottle. What we can do is have contingency plans. Revisit your BCP with these outages in mind. Start with reviewing your resiliency of these services. Consider how the work can be done if these are hard down. Have approved workarounds. Consider carefully proprietary information being processed on personal or other non-vetted platforms. As a mentor once told me: Your staff are paid to solve problems; don't be the problem they solve.

The key question is, why did "normal load testing" NOT detect the issue? Cloud-based performance testing can do great things, but sometimes unique network hardware or server configurations cannot be accurately replicated.

Testing code at production scale is very, very difficult for most companies. When you have a billion customers, it's much harder. It is understandable that Microsoft will run into issues like this, but what's less understandable is the amount of time it took to recover. I think Microsoft will have to rethink their processes and procedures, and perhaps even their architecture for more resilience and faster recovery. It's a classic case of a system intended to function as an immune system attacking and killing the host instead, just like real immune systems sometimes do.
And there you have it, a cascade of events culminating in a critical outage. Two quick points: AI-assisted code review clearly still has some homework to do, and Microsoft needs to seriously rethink its capacity limits given how much of the global economy runs on its servers. Naturally, Microsoft can expect a trip to Capitol Hill to answer for the failure.

A "bug" was not the problem. Inadequate change management was the problem.
For approximately five hours beginning around 3:30 PM EST on Friday, November 13, 2026, the US Federal Aviation Administration (FAA) ordered a nationwide ground stop of all aircraft after the FAA Telecommunications Infrastructure (FTI) experienced an unexpected volume of data that rendered some systems unstable. According to the FAA's FAQ on the incident, ongoing investigation indicates that about 1,000 connection attempts every 4-5 seconds started being broadcast from a new router that contractor L3Harris was installing as part of a "routine replacement" in the FAA's legacy time-division multiplexing (TDM) system at a site in El Dorado, Arkansas; some of the traffic also came from a radar site in Amarillo, Texas. Upon discovering the data inconsistencies, L3Harris blocked both locations and disabled the routers' broadcasting function. The FTI remained operational, but the flood of requests had caused instability in 17 out of 22 Air Route Traffic Control Centers as well as in Terminal Radar Approach Controls, prompting the FAA to implement the ground stop and wait to confirm the network was stable before resuming some air traffic around 8:15 PM. About 12,462 flights were cancelled during the ground stop. Standard procedures ensured "certain radio frequencies and radar feeds remained operational [...] to safely land commercial aircraft and ensure that critical life-saving, law enforcement and military flights continued;" about 3,500 aircraft were flying during the time of the ground stop, and all landed safely. The FAA has found no evidence of a cyberattack or malicious actor involved. Research presented in OT.today by ISMG, the CyberEdBoard, INE, and CyberEd.io analyzes the possible causes, gives a detailed timeline of the system outages, and describes related issues during L3Harris's upgrade process.

This issue is a great example that cyber resilience is not simply about protecting systems from malicious attackers. Equipment failures, issues caused by weather, configuration errors, and poorly managed changes can have consequences just as serious as a cyberattack. Organisations operating critical infrastructure need to ensure their change management processes are robust, that they understand the dependencies between systems, and that they have effective fallback arrangements in place. The cause of an outage may be different, but the impact on the people relying on those services can be very much the same.

What stands out to me about this incident is that the FAA's safety procedures worked, allowing thousands of aircraft already in flight to land safely despite a major network disruption. It illustrates how the industrial Safety, Reliability, and Performance (SRP) triad can take precedence over the information security Confidentiality, Integrity, and Availability (CIA) triad when safety is at stake. By grounding flights, the FAA deliberately sacrificed availability to preserve safety, demonstrating that keeping a system operational is sometimes less important than ensuring it operates safely.

This is going to be an uncomfortable review for L3Harris. While we can't do much to help there, we can look at our processes around critical components. Make sure that you're vetting replacement gear before it's going live, and that you know the impacts of hot swapping and failover. Have you kept your system interdependency information updated? This sounds like a lot, and it is, but you need to know to plan for and mitigate side effects and collateral damage. You may need to rearchitect your system maintenance procedures to keep impacts in scope. Don't restrict this to tabletop exercises — do real world tests. You don't want to find issues the hard way.

The FAA, like Microsoft, suffered from cascading failures that made the incident difficult to recover from. The similarities end there, however. The FAA story is more about tech debt, allowed to languish for far too long. I'm reminded of the 2023 British Library breach, where decades of tech debt came due all on the same day. The library continues to recover today, three years after their ransomware incident. Hopefully the FAA can pay down their debt more quickly.
Simply put, it lays bare just how fragile the US air traffic control system really is. Trying to overhaul a decades-old system while keeping the planes flying is a monumental tightrope walk. L3Harris shouldn't get a free pass, but they shouldn't take the fall as the sole scapegoat either.

ATC IT has been falling behind almost from the start. Bringing it into the 21st Century will be expensive. However, the problem is one that money alone cannot fix. Simply allocating money will invite more corruption while the gap will continue to widen. Modern technology can, and must, be introduced in a non-disruptive way.
FAA
OT.today
AP
Nextgov/FCW
FedScoop
Ars Technica
TechCrunch
ilvanovich writes that "AI-assisted vulnerability discovery and exploitation has changed the threat landscape such that the vulnerability remediation timeline encouraged by our disclosure policy is clearly insufficient." Until now, Project Zero has followed a 90+30 disclosure policy: 90 days for vulnerability remediation with an additional 30 days for patch rollout, known as "soak time." Silvanovich notes that "soak time is no longer meaningful in a world where attackers can almost immediately determine bug details from a patch," and that it is crucial for vendors to minimize the patch gap — i.e., the amount of time between a fix's release and users installing the update. Project Zero's new vulnerability disclosure policy, Full X, seeks to "encourage faster vulnerability remediation, encourage vendors to reduce 'patch gap' [and] allow for the policy to be automated by both Project Zero and vendors." The following is how Full X works: Project Zero will disclose vulnerabilities X days after reporting them to vendors; X will decrease from 90 days to 30 days over the next 14 months. As of Friday, November 13, the Project Zero disclosure policy is Full 90; on February 2, 2027, the policy will shift to Full 60; and on January 3, 2028, the policy will become Full 30. The majority of special case exemptions are removed from the policy, with the exception of vulnerabilities that are being actively exploited, already public vulnerabilities, and vendor requests.

Project Zero is taking an aggressive approach to the long-standing problem set behind the question I keep wrestling with: "If a fix shipped this Friday, when would it actually be running in production?" I get the spirit and intent (and the reality) of how fast attackers are working out exploits in this new era of vulnerability and exploit development. That said, the enterprise still isn't there. Project Zero's point is that the patch itself gives the bug away, so the extra 30 days weren't protecting anyone. Fair. But pushing disclosure faster, even if patches are available faster, doesn't mean patches are successfully _deployed_ faster. Can we step back for a second and think about the end-users of the products for which we're accelerating disclosure? Vendors, how much friction have you removed from the upgrade process, and can you report on how many customers are getting those updates installed within the same timeframe? If you run vulnerability operations: 1) Measure how many days it really takes you to go from vendor release to installed on your most exposed systems; 2) Ask whether your emergency change process could run at that pace routinely, without a scary headline to justify it and burnout for your teams; and 3) Ask your vendors (and insist on an answer to) what mitigations they'll publish when they can't make the deadline.

Google's decision to shorten its vulnerability disclosure timelines reflects how quickly the threat landscape is changing. This new policy is also shining a welcome light on an area that I feel is not getting enough attention, and that is the time taken for organisations to apply patches once they are available. While faster disclosure may encourage vendors to address vulnerabilities faster, it places additional pressure on organisations to deploy security updates in a timely manner. Organisations need to implement effective processes to identify affected systems, assess their exposure, and deploy fixes. There is little benefit in a vendor releasing a patch quickly if organisations take months to install it.

Given the current rate of discovery and exploitation of flaws, where AI has really changed the landscape, taking ninety-plus days to resolve reported issues is untenable. Even so, both sides have work to do. If you're reporting, you need to produce actionable, clean, accurate information, not AI Slop, and if you're maintaining systems, you need to not only move to rapid response but also look closely at a model where all discovered flaws are addressed. The hard part will be deciding if exploitability is a factor if there are reports on these you'll need to filter out. The days of “we can't test everything or address every issue” are quickly becoming nonexistent.

I find it highly unlikely that this change will move the needle. Now that vulnerability discovery involves using AI models, the chances of parallel discovery and disclosure goes up immensely. Multiple groups of people using the same models and similar harnesses are discovering the same vulnerabilities within days or weeks of each other. Shortening the deadline for disclosure narrows that window, but doesn't guarantee the defenders will win the race. The vendor/maintainer still needs to create and test a patch and roll out the patch to the public, and defenders must test and install the patch. Instead of trying to patch ever faster, defenders should instead invest time and effort into broad passive and active mitigations that will buy them the time they need to get patches put in place safely with minimal disruption to operations.

We need "good guy AI" advances that quickly analyze a patch, do rapid QA analysis, and shorten the "patch gap." Google may delay this, but no matter what, change windows will have to shift to more frequent but shorter downtime for the same monthly service level.
Natalie Silvanovich is merely pointing out the new reality we all face. While Full X is an admirable objective, waiting 14 months to tighten remediation timelines from 90 days to 30 ignores the pace of modern threats. Today's adversaries aren't taking weeks or days to reverse-engineer patches; they're doing it in hours.

In order for developers to reduce patch time, AI must be as good at remediating problems as it is at discovering them. Current models are not up to that task. Developing the models that we need should not take too long.
On October 9, Fireworks AI disclosed that it had " identified a security incident involving unauthorized use of internal Fireworks credentials." Fireworks AI detected the incident on October 6, began notifying affected customers the following day, rotated Fireworks employee API keys on October 9, and "revoked additional Fireworks user API keys and rotated service-account tokens identified in a completed credential review." Fireworks has updated the October security bulletin to add that the company has notified law enforcement about the incident and has brought in Palo Alto Networks Unit 42 to help with forensic analysis. While working with Unit 42, Fireworks has taken steps to harden its security. Fireworks writes, "The evidence indicates links to a cybercriminal threat actor group, which is financially motivated and aimed at selling tokens from unauthorized access to Large Language Models (LLMs) to illicit routing gateways." Companies affected by the incident include Uber, Cursor, and others. Uber has filed notices with Attorneys General in several US states, including California, Iowa, Massachusetts, Oregon, South Carolina, Texas, and Vermont, and has begun notifying drivers that a data breach at one of the company's tech providers (Fireworks) resulted in the theft of their personal information, including names, addresses, dates of birth, and images of their driver's licenses. The incident affects nearly 150,000 individuals. While Uber has alerts set up to notify the company of unauthorized access to their Fireworks AI environment, the alerts were not triggered, because the attackers used internal Fireworks credentials. Cursor says the Fireworks breach involved the theft of a stored access key that Cursor had issued to Fireworks, which led to "a malicious, unauthorized copy of Anthropic's Claude Code extension [appearing] in the Cursor VS Code extension marketplace."

After a breach like this, standard guidance includes "rotate your credentials." What is often overlooked is which credentials to include. In this example, organizations were compromised not just due to credentials leaked from Fireworks, but because attackers quickly used credentials they stole from Fireworks customers after gaining access with credentials exposed by Fireworks. Modern breaches move fast, and credentials need to be rotated if they were held on exposed systems. Doing so requires short-lived, centrally managed credentials that are rotated routinely.

This incident highlights how a security breach at one technology provider can have very different consequences for its customers. In this case, one customer reportedly suffered a personal data breach while another saw a compromised access key used in a software supply chain attack. What is particularly concerning is that the attackers' use of legitimate staff credentials reportedly prevented some security alerts from being triggered. Organisations need to understand what access their suppliers have to their systems and data, how that access is monitored, and what happens if the supplier's own staff credentials are compromised. Trusting a supplier should not mean automatically trusting everything done using that supplier's credentials.

Your AI gateways and AI vendors should be treated like the credential stores they are, because we continue to prove (intentionally or otherwise) that they are holding more of your secrets and data than you probably realize. The key that Cursor gave to Fireworks could only read one storage bucket, and that was still enough, because the agent traces in that bucket had credentials, leading to tentacles of access sprawling across an ecosystem. Fortunately, the fake Claude Code extension was up for about 11 hours — Cursor says "at most 5,000 users" could have received it — and Uber says it hasn't identified any misuse of driver data so far. I guess that's good news, but we're playing a game of cyber roulette, hoping that we're not an org with one of those 5,000 and that it isn't one of our dev environment admins who got it. The only guidance I can honestly give is: Get a secrets management strategy on lock. Secrets end up in traces, logs, and prompts because that's where developers put them, and you aren't going to data-scrub your way out of that. Credentials that are short-lived, tightly scoped, and kept in a vault are worth a lot less to whoever finds them later, but you need to be prepared for rapid revocation workflows on the regular these days.

As with the multi-billion dollar Binance cryptocurrency heist, it is astonishing how much damage can be done when just one or two engineers' credentials are compromised. We clearly have a lot of work to do here as an industry to prevent a single employee from being able to cause such a widespread issue. I'm hoping the ongoing trend of migrating to just-in-time, ephemeral, narrowly-scoped credentials can help here. Another case where the term "cascading" seems appropriate, as the hack of a widely-used service provider can result in a breach for hundreds or thousands of their customers, in a single day.

Using a stolen credential to distribute a malicious plugin isn't a new idea. Rotate all your Fireworks API keys and credentials. Make sure that you had the notice from them to do so; you may need to update communication preferences. Make sure that your copy of Cursor is a legitimate one, and hunt for the IoCs.

Fireworks seems to have reacted pretty quickly, but this is a good scenario for an in-depth "what if"/tabletop exercise for the scenario where your cloud service provider has a breach that enables access to your dev environment. The post-incident security hardening procedures Fireworks lists in "What We Have Done" could have prevented this incident — many dev environments are more squishy than hard.
Fireworks
Fireworks
BleepingComputer
SecurityWeek
The Hacker News
TechCrunch
SecurityWeek
OAG CA
Cursor
Researchers are finding more vulnerabilities in internet-connected refrigeration controllers. VulnCheck discovered 14 security issues in the XWEB EVO Copeland and Dixell refrigeration supervisor; XWEB EVO system is one of the systems to have been targeted in the October attacks against European supermarkets. Of those 14 vulnerabilities, two are being actively exploited: CVE-2026-124014, CVSS score 8.8, a high-severity unauthenticated arbitrary file write vulnerability in XWEB EVO 4.3.0 through 5.8.8, and CVE-2026-124010, CVSS score 9.3, a critical command injection vulnerability in XWEB EVO before 5.6.2. Patches are available for some the vulnerabilities, but because Copeland ended support for XWEB EVO in March 2026, five vulnerabilities will remain unpatched. VulnCheck writes that "XWEB EVO ships a login that runs the username as a shell command, a setuid-root interpreter that makes every page a root context, a hardcoded master password, and a daily admin credential anyone can compute from a leaked MAC, on hardware the vendor will not patch again. Operators who still run one should treat it as a device that must never touch the public internet, segment it behind a VPN, and plan the move to a supported platform." Researchers at Dutch security firm Midnight Blue have discovered six vulnerabilities in Danfoss AK-SM800A refrigeration controllers, which were also targeted in the October attacks. The flaws could be exploited to take control of devices without logging in or providing access credentials. Midnight Blue says there are currently no fixes available and that the most recent firmware release, version 5.4.0, does not address the vulnerabilities. Cybersecurity agencies in Belgium, Germany, and Ireland have urged operators to remove the devices' management interfaces from the internet. Midnight Blue has not released details about the vulnerabilities while it is working with Danfoss on disclosure plans.

What concerns me about these vulnerabilities is that some of the affected systems are no longer supported by their manufacturers, while others have no patches available. What is more concerning is that the issues are not just restricted to refrigerators but affect many other internet connected devices. Many of these systems cannot be switched off or replaced overnight. They support critical services in many sectors such as food production, retail, healthcare, transport, and communications. Organisations operating these systems need to understand which devices they have, whether they are exposed to the internet, and what compensating security controls can be put in place. More importantly, businesses need to plan to replace unsupported equipment. Just because a device is still working does not mean it is safe to continue using it. The introduction of the EU Cyber Resilience Act will help tackle this problem over time, as vendors will be required to provide security updates to their products for the lifetime of those products.

Forget about patching these Danfoss units — they belong to a class of devices I call "IT asbestos." Devices in this category are so fundamentally insecure that no amount of patches will make them secure enough to expose to untrusted networks like the public Internet. The only action to take here is to remove them from the Internet. This includes 4G/5G connected devices assigning IPv6 addresses (they will be harder to find, but just as vulnerable). Put them on a management VLAN that can only be accessed via VPN, ZTNA, or some other remote access technology. Ensure it cannot be accessed by untrusted parties.

Refrigeration controllers are the new shiny target. I'm fairly certain other OT systems are equally vulnerable and need similar protections. Remember, these are built for long life and specific purposes, not for survival on the Internet or under the current threat landscape. Don't assume they will have timely updates nor that updates can be easily or rapidly applied. With that in mind, implement appropriate protection and monitoring.

These latest findings reveal fundamental security design flaws in refrigeration controllers, including hardcoded credentials, unauthenticated remote access, and vulnerabilities that may never be patched. The reality is that insecure and unsupported equipment will remain operational for years, leaving operators dependent on protections outside the devices themselves. Network segmentation and deny-by-default firewalls can prevent unauthorized access to vulnerable controllers, while VPNs with strong authentication and restricted vendor access can reduce the risk of attackers exploiting hardcoded credentials or remote code execution flaws.

The Microsoft and FAA ground stop outages demonstrated the need for back-up procedures when telecoms are disrupted. Those same procedures are needed here — remote connectivity to the vulnerable systems needs to "have an outage" until controllers are updated.
As expected, attention has shifted toward discovering similar flaws in refrigeration systems. The report sparks an important debate regarding vendor responsibility for patching devices so soon after a recent end-of-life (EOL) notice. While policymakers and vendors debate liability, immediate user action is required: Isolate these devices from the public internet or implement strict network segmentation to prevent exploitation.

Glad someone is looking more closely into these controllers. Not so glad that it isn't the manufacturer doing this work.

Would that this problem was limited to refrigeration. The problem is the way that IT is being used to control OT. Fixing refrigeration is the wrong approach.
CyberScoop
SecurityWeek
The Hacker News
VulnCheck
Midnight Blue
Cooling Post
NCSC Ireland
Belgium
NIST
NIST
A GitHub credential theft attack began on November 10, 2026, using 4,837 compromised accounts to target over 60,000 repositories with probe pull requests, and resulting in at least 1,835 repositories running a malicious payload that exfiltrates credentials. According to Aikido Security, the threat actor uses stolen GitHub tokens to probe each repository the token can write to. If permissions and configuration allow it, the threat actor then pushes a credential stealing script as a workflow injection, install hook, or throwaway branch, and finally deletes the workflow and logs, reverting the commit within one minute of the infection; "a missing run number in the workflow's history is often the only trace." The malware itself (named "wfx") comprehensively collects secrets, variables, credentials and credential files, shell histories, "the environment and memory of running processes on the runner," and other information, before pursuing "deeper cloud access" in OIDC exchange, AWS, GCP, Azure, Vault, and Kubernetes. It takes steps for persistence and then exfiltrates a gzipped tarball of stolen data to various delivery channels depending on network egress restrictions. StepSecurity notes that about 90% of the 60-thousand-plus pull requests were still open on November 12, with new ones still appearing. Both Aikido and Step Security note that several accounts impacted by this wfx attack were also compromised by the GhostAction credential stealer in October. Both security reports provide indicators of compromise (IoCs), and affected users are urged to rotate all secrets and credentials accessible to the malware, remove its persistence, and "revoke the [compromised] account's personal access tokens, OAuth app grants and SSH keys, and sign out its sessions." While the malware has spread via some automated publishing triggers, Aikido notes that "the threat actor appears to be targeting CI secrets rather than attempting to proliferate through the supply chain."

This is a good time to sit down with your developers and have them explain how they handle pull requests, to include how they QA changes and how they would roll back a malicious update. Then downshift into verifying they are using the current best practices for managing secrets and credentials. This is about building a partnership with them. If you have a chance, make sure you're managing all your human and nonhuman identities centrally.

Another case where long-lived credentials cause chaos. Continuing the "cascading" trend, this attack functioned almost like a worm would, testing how each new compromised credential could extend the attack's reach. As an industry, we killed the idea of regularly rotating employee credentials as a bad idea, but that's because they were human-facing. Machine-managed credentials don't suffer from the same issues. The guidance from these incidents is always to rotate and revoke credentials. What if you were already doing that? All the time? The moment they were no longer needed?
Aikido
StepSecurity
Help Net Security
SecurityWeek
The Hacker News
BleepingComputer
The US Federal Bureau of Investigation (FBI) has published a Flash Alert warning that the Artex AI-powered penetration testing tool is being used to launch cyberattacks against organizations globally. Artex appears to have been used to enable the recent ConnectWise supply-chain breach. The attacks, which have targeted organizations in multiple sectors, have been ongoing since at least August 2026. The Alert notes that "ARTEX significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours." While the original developer removed the Artex GitHub repository once it became evident that Artex was being used for malicious purposes, "the framework continues to proliferate through community forks, localized variations, archived copies, and a commercial version." The document includes Artex's tactics, techniques, and procedures (TTPs) and associated indicators of compromise (IoCs), as well as suggested mitigations for "reduc[ing] the likelihood and impact of ARTEX-enabled intrusions."

Artex is just one example of a penetration testing harness used in current attacks. Things are moving fast, and it is probably not even the best anymore. Do not defend against Artex. Defend against fast-moving attacks.

Artex’s creator took down the open-source repository (it's now a closed-source repository), and backdoored versions remain. The risk with Artex is that it facilitates stringing multiple attack techniques together such as password spraying, stolen secrets, credential abuse, and lateral movement, greatly lowering the bar for an attacker. The IC3 bulletin has IoCs for your threat hunters as well as TTPs you can leverage to shape your protections.

I think most of the news coverage is focused on the wrong thing here. Artex is getting the scary-new-AI-weapon treatment (with plenty of marketing hype to go along with it), but the toolkit isn't the story any more than Cobalt Strike was when attackers picked that up. Tools built for legitimate testing get abused, and they always have. The only thing that separates a tool from a weapon is the regulation and intent around it.
What matters to you is how the risk landscape changes. The FBI put it as well as I could: it has "reduced the barrier to entry" for "threat actors with minimal expertise," and small groups running concurrent campaigns "at a scale previously unavailable to them." The techniques in the alerts and forensics analysis I've been reading are familiar ones. If you want some positive news, at least for now, the reporting around Artex-driven attacks tends to highlight how noisy they are. Anyone who's used LLM workflows can attest to the amount of “gunk” they are prone to leave on disk, and my experience with these toolsets is that they are no different. Stealth doesn't seem to be a priority for them yet, which works in the defenders' favor. If you're paying attention, you will see it, but can you act rapidly enough? That, for now, is what the threat actors are banking on. So, who's ready? 1) Build your plan around response speed, baseline deviations, and reducing scope of identity and token permissions to reduce damage — don't worry so much about IoCs and specific tools in use; 2) Go look at the peripheral systems like partner portals, legacy services, and mobile backends; and 3) Measure your own time from alert to containment, and find out what your SOC is pre-approved to do without waiting on a chain of sign-offs at 2 AM. If the honest answer is longer than an hour, you have some work to do.

There's nothing technically remarkable about what AI pentesting harnesses like Cairn and ARTEX are doing; they're doing what pentesters have been doing for decades, using the same tools and techniques. What makes it a big deal is that this ability is now available to every script kiddy, and it scales as far as attackers' token budgets will take it. This is a wakeup call that most companies have never been very secure. The ratio of predators to prey was such that the majority of poorly secured companies were able to avoid getting breached. Tools like ARTEX are changing this math.
Is anyone really surprised to see an open-source utility co-opted by threat actors? As the saying goes, once it's on the internet, it never really goes away. I'm glad the FBI put out the alert, but I doubt it will slow down the use of Artex.
The Record
The Register
SCWorld
BleepingComputer
Censys
IC3
Infoblox has confirmed that a breach of a third-party technology provider's system compromised customer and staff data. The company told news outlets that it is "investigating a security event involving unauthorized access to certain company information through a compromised third-party data migration vendor." The unnamed vendor reportedly had trusted access to Infoblox's SharePoint system, and the attackers used this access to steal the data, which "included configuration information related to some customer's NIOS (Network Identity Operating System) deployments." In a November 12 filing with the US Securities and Exchange Commission (SEC), Taiwan Semiconductor Manufacturing Company (TSMC) confirmed that data belonging to current and former employees were compromised through a breach of one of the company's cloud service providers. TSMC is notifying affected individuals and "providing them with necessary support."

Whether deliberate or accidental, make sure that you're not over-permissioning accounts. Particularly with third parties. Verify you're actively reviewing access granted, revoking rapidly when unneeded. Don't get trapped by the argument that they used to have access to that information and already know it. It's that they no longer do, and you need to protect the data owners from them or services working on your behalf, which may not all be altruistic. Watch for gaps in your logic, and accept feedback, but don't get drawn down the rabbit hole.

Only questions here, given the lack of details shared so far. Did the third-party data migration vendor still need access to Infoblox's SharePoint? Was this a well-timed attack, or did someone at Infoblox forget to revoke access once the data migration was done? We'll have to wait and see if more details are released on this incident.

In this case, better PI than IP.
The Register
TechCrunch
The Record
BleepingComputer
SEC
A cyberattack targeting business mail platform Maileva — a subsidiary of Docaposte, the digital arm of French postal service La Poste — has compromised the contact information of more than 22 million individuals. Maileva allows customers to upload documents they wish to send via post; the company then prints those documents and mails them. Maileva has more than 70,000 registered users and reportedly sends more than 180 million documents annually. A statement from Docaposte says that the French Data Protection Authority (CNIL) and affected organizations and individuals have been notified, and that an investigation is underway. Lithuania's customs agency has restored online services following a cyberattack that may have compromised personal information of hundreds of businesses and individuals. Lithuanian Customs detected the incident on November 10 and temporarily shut down the Common User Management Portal (BAP). An investigation has determined that the breach affected users who had accessed BAP on November 9 and 10. The City of Osaka, Japan has disclosed that a data breach at a third-party technology provider compromised more than one million email addresses. The "email addresses of users of the Osaka City Administrative Online System have been leaked to an external party due to unauthorized access to the email distribution service used by TKC Corporation … the provider of this system," according to a statement from the city.

Two things occur to me: First, when creating a simple service to solve problems, like the French upload/print and mail service, don't forget to double check the security and abuse paths; second, we all need to take steps to reduce the value of capture of our email or other semi-public information. This isn't just ID/credential monitoring and protection, it's also making sure that your technical protections, such as EDR, are updated and fully configured. This doesn't mean you shouldn't still use obfuscation of email, nor that you can carelessly share your information; it's that things happen and you need to stack the deck in your favor.
The Register
La Poste Groupe
Cyberattaque
The Record
LRV
LRV
The Record
Osaka
SANS Internet Storm Center StormCast Tuesday, November 17, 2026
LiteLLM/Starlette RCE Scans (CVE-2026-42271, CVE-2026-48710); Danfoss AK-SM800A RCEs; cyrus-sasl SRP Bypass (CVE-2026-129483); NoisyTCP (CVE-2026-99288)
https://isc.sans.edu/podcastdetail/10182
In this episode, Internet Storm Center honeypots are seeing LiteLLM and Starlette exploit scans, Midnight Blue found unpatched remote code execution flaws in Danfoss AK-SM800A refrigeration controllers, a Cyrus SASL Secure Remote Password bypass can allow logins without a password, and NoisyTCP is a Linux kernel TCP denial-of-service bug.
Scans for LiteLLM/Starlette (CVE-2026-42271, CVE-2026-48710)
Internet Storm Center honeypots are seeing scans that chain a LiteLLM remote code execution bug in a test URL, CVE-2026-42271, with CVE-2026-48710, a Starlette Host header validation flaw that bypasses access control. Patches are available, and Horizon3.ai has a write-up of the chain.
https://isc.sans.edu/diary/33522
https://horizon3.ai/attack-research/vulnerabilities/cve-2026-42271-chained-with-cve-2026-48710/
Danfoss AK-SM800A Remote Code Execution Vulnerabilities
Midnight Blue found six pre-authentication remote code execution flaws, including a Host header buffer overflow, in Danfoss AK-SM800A refrigeration controllers after about an hour of manual firmware review. The devices lack standard protections such as stack canaries and address space layout randomization (ASLR), no patches are available yet, and the controllers should be kept isolated from the internet.
https://www.midnightblue.nl/blog/multiple-rce-danfoss-aksm800a
cyrus-sasl SRP Server Authentication Bypass (CVE-2026-129483)
A Secure Remote Password (SRP) bug in the Cyrus SASL library, often used to authenticate Internet Message Access Protocol (IMAP) and other mail services, lets an attacker log in as an existing user with no password. SRP is usually not enabled, and mail servers that offer it typically advertise the algorithm in their banner, so a quick scan can show whether it is in use.
https://github.com/cyrusimap/cyrus-sasl/issues/896
NoisyTCP: Linux Kernel TCP Pre-auth Remote Corruption (CVE-2026-99288)
NoisyTCP is a Linux kernel Transmission Control Protocol (TCP) denial-of-service vulnerability that can be triggered against any listening TCP socket, with no indication that code execution is possible. A patch is in the kernel.org tree, exploitability varies from about 14 hours on a quiet bare-metal server to much faster on a virtual machine, and disabling SYN cookies is a possible emergency workaround but is not recommended as a proactive fix.
https://seclists.org/oss-sec/2026/q4/715
SANS Internet Storm Center StormCast Monday, November 16, 2026
Macfinger ClickFix DownTheDRAIN; Microsoft 365 Outage Phishing; Check Point PHP2Shell Exploits (CVE-2026-109618); Cursor Claude Code Extension
https://isc.sans.edu/podcastdetail/10180
This episode covers a Macfinger ClickFix campaign that delivers the DownTheDRAIN macOS infostealer, phishing that leverages last week's Microsoft 365 outage, in-the-wild Check Point Gateway PHP2Shell attacks, and Cursor's post-mortem on a compromised Anthropic Claude Code extension.
Macfinger ClickFix Campaign Delivers DownTheDRAIN Infostealer
Brad Duncan followed up on his September diaries about the Macfinger ClickFix campaign, which tricks macOS users into pasting a command into Terminal and then loads additional malware. Huntress also wrote about the payload, which Brad nicknames DownTheDRAIN, a cryptocurrency wallet drainer; the sample he tested was blocked by Apple.
https://isc.sans.edu/diary/33518
https://www.huntress.com/blog/mac-crypto-draining-malware
Microsoft 365 Outage Phishing and a Failed Pool Alarm
Microsoft Threat Intelligence says attackers are sending phishing emails that leverage last week's Microsoft 365 outage, claiming backed-up or queued mail may be deleted if the recipient does not respond. In Germany, a chlorine-level alarm email for a swimming pool never arrived during the outage, and several people were injured at a Phantasialand hotel pool; Johannes notes that life-safety systems should not depend on a single email service.
https://x.com/MsftSecIntel/status/2121491564161581148
Check Point Gateway PHP2Shell Attacks
Check Point Research documented attacks against Check Point Security Gateways that began in late October, about a week after patches were released, exploiting the PHP2Shell remote code execution vulnerability (CVE-2026-109618 and CVE-2026-109619). The attackers also used a privilege escalation vulnerability in the DynamicCLI interface (CVE-2026-123039) and dropped a web shell, then a proxy, a wiper, and other components; Johannes says checking devices and web applications for web shells is a good generic test for compromise.
https://research.checkpoint.com/2026/php2shell-exploitation-on-check-point-gateway/
Cursor Post-Mortem on the Claude Code Extension Compromise
Cursor published a post-mortem on the compromise of the official Anthropic Claude Code extension in the Cursor marketplace. The intrusion started with a credential exposed at Fireworks.ai that led to customer communications containing additional credentials; monitoring identified the unauthorized activity and keys were rotated, but one missed credential was then used to compromise the extension.
https://cursor.com/blog/security-incident-update
My Upcoming Classes
Catch up on recent editions of NewsBites or browse our full archive of expert-curated cybersecurity news.
Find Control Gaps Before the Next Audit Does. DataBee continuously measures security controls against data from the tools you already use, giving you current, traceable evidence of what’s working and what isn’t. Test it against one of your own priority controls,
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