The FBI's Friday flash alert on Artex describes an open-source framework that takes any large language model and turns it into an attack system. It was published on GitHub in August by an account called Autumn-27. Its developer, Chinese security engineer Li Puhua, said on 8 October that it had been abused by malicious actors, denied responsibility, and deleted the repository. Removal did not work. The FBI says the framework continues to spread through forks, localised variants, archived copies and a commercial edition.
The Bureau reports intrusions reaching privileged access from initial targeting in under three hours, and says the framework lets operators with minimal skill run complex campaigns against many targets at once. Censys counted over 1,600 publicly exposed consoles, with anything behind a firewall uncounted. Attacker instances ran mostly on Chinese models, the DeepSeek Flash family favoured for capability, price and, in the FBI's words, the absence of observable guardrails when the activity is framed as penetration testing.
Amazon's CISO added the detail that should stop you. His team documented Chinese-nexus groups using Artex in a self-funding loop: steal credentials that reach AI models, then spend the stolen access on inference to hunt more credentials. One campaign used it to defeat crypto exchange identity checks with deepfake video.
That is this week's frame. AI no longer merely makes attacks cheaper. It lets an operator with modest skill run attacks faster than a defender can schedule the meeting to discuss them.
The Vendor Update Server Is the Crown Jewel
Three incidents share a shape this week, and it is not the shape the vendors described.
Quantum Systems, a German defence manufacturer, confirmed on 13 November that it detected an intrusion in late October by hostile actors. Der Spiegel and the Financial Times, citing internal sources, reported that the entry was a malicious update file for QBase, the ground control and mission planning software. Whoever owns the update server for mission planning software owns the aircraft.
TSMC's case is smaller in consequence and identical in shape. On 8 November a threat actor using the alias SmithBianl posted a sample from what they claimed were 6.9GB of employee data covering 46,000 records, including national ID numbers, addresses, phone numbers and bank details. Bidding started at 3 BTC, about 65,000 dollars, with free publication threatened by 15 November. TSMC told the SEC that the material matched data held by a cloud service provider, that it found no evidence of compromise in its own systems, and that operations and customers are unaffected. All of that may be accurate. It also confirms that employees' ID numbers sat in someone else's storage, under someone else's controls, and are now for sale.
Cursor completes the pattern, and is the most instructive of the three. Attackers took an access key held at Fireworks AI, a third-party provider Cursor used for model training and inference, and read a bucket of historical coding-agent traces. Some traces contained unredacted API credentials. Those credentials opened further Cursor systems, and from there the attackers reached a key that could publish to the company's extension marketplace. On 29 October they shipped malicious builds of Anthropic's Claude Code extension, exposing up to 5,000 developers to malware that installed persistent SSH backdoors. Cursor's figures (fewer than 0.05% of users with traces exposed, fewer than 0.01% of customers with credentials compromised) are true and beside the point. None of those figures matters once a poisoned extension sits in an IDE.
The lesson is not third-party risk in the abstract. A vendor's authentication perimeter means nothing if the vendor handed a subcontractor read access to a bucket with keys in it.
End of Life Is a Date, Not a Project
Microsoft reminded the world on 16 November, for the second time in a month, that Windows Server 2016 leaves extended support on 13 January 2027. The January 2027 security update will be the last. Patches then continue only for machines on Extended Security Updates. Windows Server 2016 Essentials features end on the same day.
Sixty days is not a migration window. It is enough time to buy ESU, identify stranded workloads, and tell the board which you are doing. Drifting into the date produces the one outcome nobody defends a year later: an unpatched system with no programme behind it.
Linux produced the sharper reminder. NoisyTCP, disclosed on 16 November as CVE-2026-99288, is an 11-year-old flaw in the kernel's TCP stack. A NULL pointer dereference lets an unauthenticated attacker crash a host running Nginx, Apache or an SSH server in default configuration, using crafted packets that exploit the timing between SYN cookie handling and socket hashing. A stale hash lookup must race the freeing and reallocation of a request socket across receive contexts, and packet processing must spread across two or more receive queues.
The mitigating facts are real. The result is a kernel panic, not code execution, the race window is narrow, and the discoverer notes a quiet bare-metal server may take a long time to fall. So NoisyTCP is a poor weapon and a good nuisance: the hosts that matter are the ones an adversary wants kept offline. The patch is available and the fix entered mainline two weeks before disclosure. Where you cannot apply it, the interim advice is to disable SYN cookies and serialise to a single receive queue, which buys the mitigation by surrendering SYN flood protection.
CISA's 13 November additions to the Known Exploited Vulnerabilities catalog included CVE-2026-114713 in UnrealIRCd and CVE-2026-123039 in Check Point products. The second is the permission assignment flaw exploited against gateways last week. Its KEV entry confirms rather than discovers, and is the last argument anyone needs with a team that has not applied the November hotfix.
The Regulator Found Its Line
Ofcom published finalised child safety rules on 16 November and said plainly that it lacks the power under the Online Safety Act to regulate devices, which closes the immediate question of client-side scanning. Campaigners wanted on-device detection of child abuse material in encrypted messages; researchers and privacy groups argued that scanning on the device breaks encryption. Apple tried it in 2021 and shelved the plan in 2022. Ofcom left a door open through Section 121, which allows a technology notice compelling accredited software to find such material in private messages, but no tool is accredited and nothing has been announced.
The parts that affect your operations are less discussed. Platforms must detect known and unfamiliar abuse material in public content, may judge an encrypted group from its profile picture, description or coded terms backed by a police report, and must permanently ban adults found sharing such material or grooming children. Ofcom conceded this may interfere significantly with privacy. The rules take effect in February 2027. Ofcom also dropped planned fraud detection requirements.
Encrypted group membership is now something that can be inferred from metadata, which makes it worth a line in your acceptable use guidance. And a regulator who states in writing that it lacks a power tends to come back for it. Section 121 will return when accreditation exists.
The House of Lords supplied the political counterpoint the same day. The National Resilience Committee reported that the UK is underprepared for potentially catastrophic risks from Russia, China and Iran, and called the government complacent. Its report arrives eight days before a major vote on the Cyber Security and Resilience Bill, five years in development, whose scope it calls limited and which leaves most of the public sector outside. Peers had already defeated the government twice in October on amendments for periodic reviews and a register of hostile states. Direction is clear: the UK's regulatory surface will widen and enforcement will harden.
Smaller Bites Worth Your Attention
Fandango disclosed that skimmer code on its checkout pages captured card numbers, CVVs, expiry dates, billing postcodes and email addresses between 8 and 11 October. It established on 30 October that customers were affected, and offers two years of monitoring. No victim count was published. A four-day dwell time on a payment page serving 31,000 US cinema screens is a web application problem you can test for.
Zimperium's zLabs documented ArcticRAT feeding stolen SMS and installed app lists to xAI's Grok, identifying banking and crypto applications, filtering one-time passwords and transfer alerts, and estimating the victim's financial standing. Zimperium notes, correctly, that the analysis works on data already stolen and does not reach the account. The malware drives banking apps through Android accessibility services, and can paint a black overlay so the handset looks powered off while an attacker works. It also shows a fake ransomware screen that encrypts nothing, because frightening people is cheap.
Microsoft's preliminary report on the 12 November Microsoft 365 outage deserves reading for reasons that are not about Microsoft. A bug introduced in July in a networking gateway update sat unnoticed for over three months. Load testing missed it because lab hardware had fewer cores, and the fault worsens with core count. Teams failed over onto Exchange Online's networking infrastructure about 15 hours before the incident, raising traffic, while almost 14% of gateway capacity was already offline for an unrelated reason. Within two hours the defect had spread across the fleet, and more than 80% of servers were unresponsive. Regional isolation safeguards did not contain it. Service was restored at 11:10 ET on 13 November.
Any continuity plan assuming email survives a Microsoft incident has just been tested by the vendor, and failed.
What To Do This Week
Demand an answer on Artex exposure. Censys counted 1,600 exposed consoles and the FBI notes the framework's preference for partner portals, mobile backends, employee applications, sales support systems and applications built outside central IT. Sweep the internet-facing estate nobody owns, then check access logs for credential hunting against AI provider keys, the self-funding loop Amazon described.
Audit every update path with system rights. Quantum Systems is the case study. Your equivalents are backup agents, EDR connectors, hypervisor tools, firmware updaters and OT engineering workstations. Count them, assign owners, and confirm the vendor signs and you verify.
Treat marketplaces as supply chain. Cursor's breach ended in a poisoned extension inside an IDE, delivered by the vendor's own marketplace. Know who can publish into what your developers consume, and whether they passed any review you control.
Hunt for secrets in anything you have sent a model. Historical traces, prompts, support tickets and transcripts all accumulate in third-party systems. Query for key formats, revoke what you find, and write the rule that stops the next batch.
Decide the Server 2016 question in writing this month. Either migration completes by 13 January 2027 or you are buying Extended Security Updates. Both are defensible, and neither happens by default.
Make the NoisyTCP trade deliberately. Patch, or accept the SYN cookie and receive queue changes, and record why for every host where you choose the mitigation.
The Artex advisory will be cited as the point where machine-speed attack moved from projection to documentation. What makes it hard is not the capability. Every standard control assumes a pace nobody keeps now. The organisations that come through the next year in decent shape will not have the most tooling. They will have update paths, credentials and trusted channels few enough to enumerate, because enumeration is the only defence that scales against an attacker able to try everything at once.

