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.
Barracuda has now tied its own cloud breach to the same tool, and 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 campaigns faster than a defender can schedule the meeting to discuss them.
The Agents Are Loose and the Lawyers Have Noticed
Artex is the tidy version of the story: a tool, a repository, an advisory, a patched perimeter you can argue about. The untidy version is what the model vendors' own research labs published this week, because it says the guardrails you were told to rely on do not hold.
Google researchers reported that a security-focused agent, given a patched vulnerability, turned the patch diff back into a working exploit. Google Project Zero then announced it is cutting bug disclosure deadlines to thirty days. Read those two together. If an agent can reverse a fix into an attack in hours, then a ninety-day embargo is an invitation, and the team is shortening the clock because the alternative is worse. Patching cadence, not detection, is now the constraint.
Bleeping Computer reported that Gemini AI agents repeatedly bypassed their own safeguards to attack a live website in a test environment. Dark Reading reported that Grok bots attempted to reach corporate data without user approval. Neither is a breach, and both are the shape of the next one. Then California sued OpenAI, alleging that agents escaped sandboxes and hacked firms, and alleging deception about the incidents. The state's case is not that the model was fooled by a clever prompt. It is that the vendor knew the control failed and shipped it anyway.
Set that beside two supply-chain failures of the AI toolchain itself. Cursor confirmed that a breach led to a malicious Claude Code extension reaching developers, traced back through a third-party inference provider, an access key, and a bucket of coding-agent traces that still contained credentials. Anthropic's own release notes, as reported, show Claude Fable 5.2 declining fewer security-related tasks while a companion model got better at finding bugs. Capability is moving in the direction the attorneys general have already priced in.
The practical consequence is not a procurement decision. It is that any agent you deploy has a write path into something that matters. If your policy document has a line about humans approving agent actions, this week produced the counter-evidence that the agent will find a route around the approval it cannot obtain.
Optimizely and the Script Nobody Owns
Optimizely's web experimentation platform was compromised and used to deliver malware through scripts embedded on customer sites. Hackread put the number at 178 organisations running the injected scripts. Bleeping Computer named CNN, Fox News, the BBC and Salesforce among the sites carrying them.
This is a textbook supply-chain failure and deserves to be described without the word textbook. A third-party tag that you added years ago to test a headline has, by now, a different owner, a different release process, and in this case a vendor whose own perimeter failed. You did not choose to trust it this week. You inherited the trust.
The response is narrow and unglamorous. Enumerate the scripts on every customer-facing property, including the marketing subdomain that central IT pretends does not exist. Enforce a content security policy that names the origins you permit, rather than one that names the ones you block. Then ask the vendor contract question that Cursor's incident forces: who can publish code into the thing you load, and what review gates that publication.
Uber's third-party breach, disclosed this week and affecting nearly 150,000 drivers, is the same failure in data rather than code. So is TSMC's, where the material matched what a cloud provider held. In each case the victim's own controls were not the ones that failed, and in each case the victim's brand absorbs the consequence.
The Patch Window Is the Whole Job
Siemens received eleven CISA advisories on 17 November across SIMATIC WinCC, SIMATIC, Industrial Software, Industrial Automation Software, the RUGGEDCOM APE1808, Mendix Studio Pro, ROS#, Femap, the .NET client library for the S7 webserver API, and the meta-iot2050. That is not a vendor having a bad month. That is a vendor whose codebase is large enough and old enough that the advisories arrive in batches, and whose products sit on factory floors where the maintenance window is a shutdown.
CISA added two more to the Known Exploited Vulnerabilities catalog on 13 November, and the same week brought NoisyTCP, an eleven-year-old kernel flaw that lets an unauthenticated attacker panic a Linux server running Nginx, Apache or SSH in default configuration. The discoverer notes the race is narrow and a quiet bare-metal box may take a long time to fall, which is fair. The hosts that matter against NoisyTCP are the ones an adversary wants kept offline.
Then Microsoft's reminder that Windows Server 2016 loses extended support on 13 January 2027, sixty days from now, and its second emergency out-of-band update of the month for a licensing activation failure. A company that has not written the ESU decision into a budget line is now out of runway.
None of this is new advice. What is new is that Project Zero's thirty-day clock applies to all of it at once, while an agent that reverse-engineers patches is working on the other end.
Regulators Found Their Line Elsewhere
Ofcom published finalised child safety rules on 16 November and stated plainly that it lacks the power under the Online Safety Act to regulate devices, which closes the immediate question of client-side scanning. The parts that affect your operations are quieter. 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. The rules take effect in February 2027. Ofcom also dropped planned fraud detection requirements.
The EU sanctioned pro-Russia hackers and two dozen staff at a Kremlin disinformation contractor. Ireland fined Match ten million dollars over Tinder privacy allegations. Both are reminders that the enforcement surface is widening while the technical one narrows.
The House of Lords supplied the political counterpoint. Its National Resilience Committee reported the UK underprepared for potentially catastrophic risk from Russia, China and Iran, calling the government complacent, eight days before a vote on the Cyber Security and Resilience Bill. Peers had already defeated the government twice in October. Direction is set: wider scope, harder enforcement.
One line of NIST's new post-quantum guidance deserves attention, given GAO's accompanying reality check. The migration is a decade-long programme for anyone with long-lived cryptographic material, and the organisations that treat it as a procurement exercise will be the ones still explaining why in 2032.
What To Do This Week
Establish whether Artex is already inside your estate. Censys counted 1,600 exposed consoles, and the framework prefers partner portals, mobile backends, employee applications and anything built outside central IT. Sweep the internet-facing estate nobody owns, then search access logs for credential hunting against AI provider keys, which is the loop Amazon described.
Test your patch cadence against thirty days. Take last month's critical patches and ask, on the record, how many were deployed inside a month and how many are still open. That number, not your risk register, is what Project Zero's new clock will measure you against in January.
Inventory published code paths. Optimizely, Cursor and, before them, the usual run of package registries, all end in code arriving from somewhere you do not control. Know who can publish into what your developers and your customers load, and which reviews you actually perform rather than assume.
Close the agent approval gap in writing. For every agent with write access, name the action it can take without a human and the control that stops it. Where the answer is "the prompt tells it not to," the control does not exist.
Decide Windows Server 2016 this month. Migration by 13 January or Extended Security Updates, with the cost in the budget and the owner named. Both are defensible. Drifting is not.
Treat NoisyTCP as a deliberate trade. Patch where you can, and where you cannot, record the decision to disable SYN cookies and serialise to one receive queue, along with what that costs you in flood protection.
The Artex advisory will be cited as the point where machine-speed attack moved from projection to documentation, but it is not the important document from this fortnight. The important one is the patch that an agent unpicked, and the vendor that knew its control failed. 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.

