First up, we have a technical segment on how to setup Waydroid and Mitmproxy for good and evil, then in the security news:
In the security news this week:
- ARTEX: pentesting goes AI
- If everything’s urgent, nothing is
- Disclosure clocks just got shorter
- NGINX, Apache and Java, meet RCE
- Who’s actually hacking with AI?
- Refrigeration with a root problem
- Old USB, new infections
- Router trouble grounds flights
- Fridge controllers need better defenses
- SANSAR uses your GPU and hides
- Vulnerability volume keeps climbing
- AI puts defenders to the test
- Taiwanese calls, intercepted
- Air traffic gets an AI copilot
- ESP32 wants a desktop
- Script tags, now serving skimmers
- Japan’s supercomputer goes offline
- Surveillance, courtesy of ad tech
- Cloud hosts peek inside protected VMs
- LLMs can edit their own brains
- France puts DNS on trial
- Leaked keys vanish in seconds
- Flipper One levels up
The cloud isn't somebody else's computer anymore. It's yours... and so are the misconfigurations.
Attackers don't care whether it's AWS, Azure, or GCP. They care that somebody left an identity over-permissioned, a storage bucket exposed, or a Kubernetes cluster wide open.
If cloud security is part of your day job, come hang out at the Cloud Security Virtual Summit. You'll hear from practitioners sharing what's actually working to defend modern cloud environments.
Security Weekly listeners can register for free at https://securityweekly.com/cloudsec using the discount code CSS26-SW!
Paul Asadoorian
- Threat Researcher’s Notebook: Detection Strategies for ARTEX and Related Forks
Summary: ReversingLabs mapped 11 ARTEX fork families by analyzing 184 operator binaries and 3,147 unique IP indicators. A shared favicon helped identify nearly 1,000 live consoles, but forks can change their names and fingerprints, so the report emphasizes behavioral detection and hunting in network and authentication logs.
Paul's take: The useful point here is that ARTEX detection should not depend on finding one binary, hostname, or IP address. The operators can run the framework from their own infrastructure and hide behind VPNs and proxies, while forks can rebrand the console; ReversingLabs instead points to patterns such as bursts of automated scanning and authentication abuse. The underlying techniques are familiar red-team tradecraft. The change is that an LLM can chain them together quickly, which leaves less time to respond when basic controls like patching, credential hygiene, and restricting public management interfaces are missing. (ReversingLabs' analysis)
- Java4Shell, NGINX2Shell and more: Why we are not at “Yellow”
Summary: Johannes Ullrich explains why the SANS Internet Storm Center stayed at Green amid major outages and critical vulnerabilities: its Java honeypot was seeing attacks, but none exploiting Java4Shell at the time of writing. His advice is to patch, automate routine work, reduce exposure, and avoid dropping everything for each new “world-ending” threat.
Paul's take: I like this because “urgent” can’t be the default setting. Patch serious vulnerabilities, but don’t burn out a team by turning every new disclosure into a fire drill. Automate the repeatable work, reduce what’s exposed, and use AI where it helps with the jobs you struggle to keep organized. If every week is Yellow, nobody can stay on alert. (SANS ISC diary)
- Anthropic Cuts Disclosure Deadline to 30 Days For Serious Bugs
Summary: Anthropic is shortening its coordinated disclosure window for vulnerabilities it reports to vendors and maintainers: high- and critical-severity issues move from 90 days to 30, and other issues to 45. The new timelines apply to reports made on or after January 1, 2027; the seven-day window for flaws under active exploitation stays the same.
Paul's take: Anthropic’s argument is that AI makes rediscovery more likely, so a 90-day period of secrecy may no longer be realistic. That makes sense, but a shorter clock also puts more pressure on maintainers, especially small open-source teams that need time to reproduce a flaw, test a fix, and ship it. Anthropic says it will provide actionable reports and work out timelines for hard-to-update systems or flaws that require specification changes. This is Anthropic’s policy for its own vulnerability reports, not a new industry-wide deadline. (Anthropic’s policy update, Infosecurity Magazine’s report)
- seishuku/FortiGate_Firmware: Custom firmware for Fortinet Fortigate 60 router/firewall
- Rapid7
- sfewer-r7/CVE-2026-123007: Proof-of-concept (PoC) scripts for CVE-2026-123007, an integer underflow in NGINX HTTP/2.
- Who’s Hacking With AI, and What They’re Using
Summary: Daniel Miessler introduces two trackers: one cataloging 286 reported cyber incidents since 2023 where AI was used or credibly suspected, and another listing 154 open-source offensive AI harnesses. The incident tracker says 78 of 174 reported attacks in 2026 so far describe AI running part of the attack, compared with none of 27 reports from 2024; 14 harnesses have appeared in real operations, with seven tied to confirmed breaches.
Paul's take: I like the attempt to move this conversation from anecdotes to evidence, especially the links between specific tools and reported operations. But these are curated public reports, not a census of all attacks, and the 2026 numbers are a snapshot that depends on what researchers have published and how the tracker classifies AI’s role. That makes the sites useful for tracking the evidence, as long as we don’t mistake the totals for a complete measure of how often attackers use AI. (Miessler’s write-up, incident tracker, offensive harness index)
- Unauthenticated Remote Code Execution in NGINX: CVE-2026-123633 and CVE-2026-123007
- Java4Shell: Write once, RCE everywhere
- Code of the Day: Unauthenticated Root in Copeland XWEB EVO
- Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents
- Supply Chain Risk from Backdoored LoRA Models
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
- The Retadup worm revives from old USB drives to download CoffeeLoader
Summary: Copies of the Retadup worm survived the 2019 takedown on USB drives and offline computers. After someone re-registered Retadup’s old command-and-control domains, those dormant copies could receive new instructions, including to download CoffeeLoader, which can establish persistence and fetch additional malware.
Paul's take: The interesting part is that taking down a botnet does not recall the copies already sitting on removable media. When its old command-and-control domains came back to life, Retadup became a delivery mechanism for a much more modern loader. Kaspersky says it blocked about 15,000 Retadup infection attempts in the first nine months of 2026, but that is its own telemetry, not 15,000 confirmed infections or a count of the entire botnet. Kaspersky also says it cannot tell whether the person who revived the domains was Retadup’s original operator or someone else. (Securelist's technical analysis)
- Two HPE Juniper network routers grounded all US flights last Friday
Summary: After a new HPE Juniper router was installed at an FAA site in El Dorado, Arkansas, the FAA telecommunications network was flooded with what officials described as about 1,000 connection attempts every four to five seconds. The resulting instability affected communications and radar feeds at 17 of 22 air traffic control centers, prompting a nationwide ground stop; the FAA disabled a function on the Juniper routers and paused further installations.
Paul's take: I would not call this a broadcast storm yet. The FAA has used the phrase “broadcasting function,” but it has not published the protocol-level details needed to tell whether this was Ethernet broadcast traffic, multicast, routing or control-plane traffic, or something else. “Pinging” is not enough to identify the traffic, either. Network World quotes an unnamed source saying a multicast feature was enabled and normally off by default, but that has not been confirmed as the root cause. A network engineer's CellStream analysis also says the public evidence does not establish a traditional Layer 2 broadcast storm. As of November 18, the FAA's account describes the traffic surge and the steps it took, while AP reports that investigators still have not determined the root cause. HPE says the router worked as configured, so the key question is what feature and configuration caused the new equipment and the existing network to behave this way. (Network World’s reporting)
- Don’t bother patching: multiple remote code execution vulnerabilities in Danfoss AK-SM800A
Summary: Midnight Blue found six pre-authentication remote code execution flaws in the Danfoss AK-SM800A, including in firmware 5.3.2. Because the vulnerable application runs as root, a successful exploit could give an attacker control of the controller. Midnight Blue recommends restricting remote access; the full technical details are being withheld during disclosure.
Paul's take: There is now an official attribution for the October 11–12 attack in France: a French interagency cyber center that includes ANSSI, military cyber command, and France’s intelligence services attributed it to Shahid Kaveh, a unit linked to Iran’s Revolutionary Guard. Its public rationale includes shared infrastructure used in reconnaissance against Schneider Electric equipment, use of the FOFA scanning tool, amplification by a channel linked to the unit, and traces of Farsi in the attack infrastructure. (French C4 attribution report)
That does not establish that Iran was behind every later refrigerator incident, or that these six newly disclosed bugs were used in the French attack. Belgium’s interior minister said Iran was very likely behind the November incidents there, but its foreign minister said the available information was insufficient to formally blame Iran and that the investigation was ongoing. (Reuters report on the Belgian ministers' differing assessments) France’s October incident has a formal, publicly explained attribution; the November cases remain less settled. The Danfoss vulnerabilities matter either way, but we should not present them as the attack method unless investigators establish that link.
- Linux cryptomining malware: Inside the SANSAR toolkit
Summary: Elastic analyzed SANSAR, a Linux cryptomining toolkit used after an attacker gained root through an internet-exposed Ray job agent that accepted unauthenticated submissions. Across three Linux GPU hosts, the toolkit managed CPU and GPU miners over Tor, hid activity with a kernel rootkit, falsified NVIDIA utilization readings to show 0%, and included a separate backdoor that fetched its command server address from an Ethereum contract. (Elastic Security Labs)
Paul’s take: The 0% GPU reading is a neat bit of misdirection, but the initial foothold is the bigger lesson: the Ray job agent was reachable from the internet and let anyone submit work. This is cryptojacking against GPU infrastructure, including machines that may be used for AI workloads; Elastic doesn’t describe it as an attack on AI models. Once the attacker had root, they layered on Tor, a kernel rootkit, and blockchain-based C2 discovery, so checking GPU dashboards alone won’t tell you what’s running. (Elastic Security Labs)
- When AI Finds the Bugs: Rethinking Exposure in a Flood of Vulnerabilities
Summary: Akamai argues that vulnerability discovery and exploit development are accelerating, making scan-rank-patch cycles harder to keep up with. It cites roughly 300 CVEs a day in 2026 and a six-day median from publication to confirmed exploitation for flaws later added to CISA’s KEV catalog within a year. The post then pitches Akamai’s ExAR and Guardicore Segmentation as tools to identify reachable vulnerable systems and contain them while patches are pending. (Akamai)
Paul’s take: The practical point is solid: prioritize vulnerabilities by what’s actually exposed and reachable, then contain the attack path while patching catches up. But that six-day median applies only to vulnerabilities eventually added to KEV, not every disclosed flaw, and the post doesn’t establish how much of the CVE surge is caused by AI. It’s also clearly a product pitch for ExAR, and I do not believe one product is going to help you, you need AI tying together multiple tools and products, custom to your environment to be successful. (Akamai)
- GreyNoise AI Cyber Index: Benchmarking AI Models on Defensive Capabilities
Summary: GreyNoise’s AI Cyber Index evaluates 46 models on three defensive tasks: writing Suricata network-detection rules, writing YARA file-detection rules, and tracing an alert to the exploit behind it. Its 145 problems use held-back attack variants, benign lookalikes, and unrelated data to test whether answers generalize without excessive false positives. No model scored perfectly; Grok 4.5 led the refusal-adjusted composite at 71.7%. (GreyNoise)
Paul’s take: I love how Opus 5.5 scores highest in refusals, its so annoying, and others share my feelings on it too, how are we supposed to use it defend our systems if its constantly refusing to do work forus? But “defensive capabilities” is a big label for three specific tasks, and the composite depends on GreyNoise’s chosen mix and weights, with network detection counting for over half the score. Useful data for comparing models on these jobs; not proof that the top-ranked model can run your SOC. (GreyNoise)
- Critical RCE Vulnerability in Apache HTTP Server: Everything You Need To Know
Summary: CVE-2026-119618 is a race condition in Apache’s HTTP/2 module that can corrupt heap memory and allow unauthenticated remote code execution. It affects Apache HTTP Server 2.4.55 through 2.4.70 when HTTP/2 is enabled; Apache fixed it in 2.4.71. Wiz estimates 64% of cloud environments contain vulnerable resources, but notes those figures don’t account for configuration. (Wiz, Apache advisory)
Paul’s take: The HTTP/2 condition is important: Wiz’s prevalence numbers are a reason to check, not proof that 64% of cloud environments are exploitable. Also, the standalone
mod_h2version table is inconsistent: it lists 2.0.42 as both affected and patched, while Apache lists versions through 2.0.42 as affected. Check the Apache advisory or your distribution’s package notice, and patch; disabling HTTP/2 is a temporary workaround if you can’t update yet. Also important to check your devices that use Apache, various devices use a wide array of web servers. Though I am finding that devices typically do not use HTTP/2, mileage varies. (Wiz, Apache advisory)
Larry Pesce
- Hacked Asian Telcos Used to Intercept Taiwanese Calls and Texts
- Critical Apache HTTP/2 Flaw Could Let Unauthenticated Attackers Execute Code Remotely – Slashdot
- Shipping giant Maersk says recent cyber incident affected legacy system
- FAA Expands AI Advice to Manage Air Traffic Congestion Across Entire US Airspace – Slashdot
- A Desktop OS For The ESP32-P4
- Hijacked Optimizely Scripts Served Card Skimmer and iOS Exploit Kits to CNN, BBC and Zoom Visitors
- Japan takes Fugaku supercomputer offlline after cyberattack
- Custom Firmware For Fortinet FortiGate 60 Firewall/Router
- FBI warns about ARTEX
- A vulnpocalypse bug worthy of attention: Researchers disclose critical flaw in Java with help of AI
- Java4Shell
- NGinx2Shell
- Holiday Hack Cybersecurity Challenge 2026
- Everything4Shell — CVSS 11.0 / 10
Sam Bowne
- Security Agencies Worldwide Found to Use Ad-Based Surveillance System
Webloc is a global geolocation surveillance system that monitors hundreds of millions of people based on data purchased from consumer apps and digital advertising. U.S. customers include ICE, the U.S. military, and several police departments. In March 2026, 72 senators and representatives in the U.S. Congress called for an investigation into “warrantless purchases of Americans’ location data” by ICE and other U.S. agencies.
- New Mior and Skyfall Attacks Let Rogue Cloud Hosts Steal Secrets From AMD SEV-SNP VMs
SEV-SNP brings confidential computing to AMD EPYC server processors. It uses hardware-enforced memory encryption and access control to protect a VM, even from the host's hypervisor. But two attacks made public this week allow a malicious host to read data inside AMD SEV-SNP-protected virtual machines (VMs), which are built to keep the server's operator out. Updated firmware that mitigates the flaw is already out.
- LLMs can natively read & write neural network weights
This is a breakthrough! Current LLMs are trained using gradient descent, taking hundreds of millions of computation steps, and producing neural nets that are very difficult to understand. No one really knows what each neuron is doing. This paper presents a new way to assign weights to neurons--by using an LLM to analyze the neural net and understand what the functions of the neurons are, and then assigning weights to them to produce the desired output without using gradient descent. This is called Weight-Space Programming: programming a LLM as if it were Python code. The authors used this technique to improve the math performance of a trained LLM from 7.9% to 48.1% without any additional cycles of training. This way of programming a LLM may vastly improve the efficency of training.
- FBI FLASH: Threat Overview of ARTEX Autonomous Penetration Testing Framework
FBI FLASH: Threat Overview of ARTEX Autonomous Penetration Testing Framework The FBI warns us that ARTEX-powered attacks will be faster and far more capable than earlier ones. They list tactics used by current attackers and IOCs.
- Quad9 to withdraw its public DNS resolver from France after court penalties
France demanded that it block domains tied to alleged copyright infringement. Quad9 plans to stop serving French users, just as Cisco's OpenDNS did.
- Quality Isn’t The Main Worry About AI Pentesting — Data Privacy and Production Safety Are
Time-to-Exploit is now down to a single day, so an annual pen test seems pointless. 90% of security leaders say manual pen-testing is dead. Most organizations already use AI pentesting across their environment, and only 1% of leaders don't plan to use it soon. The barriers to AI pentesting include data privacy (40%), safety on production systems (36%), and whether auditors and regulators will accept AI pentest reports (28%).
- Cloud keys leaked in Docker Hub images are stolen within seconds
Researchers published 300 honeypot images across eight public container registries. The fastest observed use of a planted AWS key came 7.86 seconds after the image was published on Docker Hub. The median time for the first API call was 51.19 seconds, while all 100 AWS keys were used within 15 minutes.
- CISA Urges Immediate Patching of Java4Shell Flaw That Could Allow Remote Code Execution
"This critical memory-corruption vulnerability, known as 'Java4Shell,' affects the Base64 decoder of the x86-64 HotSpot Java Virtual Machine (JVM), which is part of the Java Development Kit (JDK)," CISA said. "Specially crafted Base64 input could crash the JVM and allow for unauthenticated remote code execution."
- Flipper Devices launches $399 ‘Flipper One’ Linux-based multi-tool
I didn't see any exciting list of new features, so I assume it has a teleporter, a time machine, and a death ray. I bet Paul knows all about it and already ordered one.
