Saturday, 3 October 2026
Thirty-Two Steps, Nobody Driving
Microsoft’s Annual Threat Report Says the Advantage Has Moved to the Attackers, and Names the Two Models That Proved It
Microsoft published its 2026 Digital Defense Report on Thursday, and the number everyone is repeating is thirty-two. In a controlled evaluation against an emulated enterprise with no defenders present, Anthropic’s Mythos and OpenAI’s GPT-5.5 each “took control of the whole domain, including the main server and all user accounts, through a 32-step attack chain.” Microsoft’s own framing of the trajectory is the part worth keeping: threat actors are moving “from AI assisting human operators, to AI directing attack activity, toward autonomous execution.” Two distinctions matter here and the report observes both. The 32-step chain is a lab exercise with the defenders switched off — it is capability, not an incident. The incident is JADEPUFFER, early July 2026, which the report carries as the first documented fully automated ransomware-extortion attack, named by Sysdig’s threat research team. The supporting figures are the ones that will actually ruin somebody’s quarter: median time from a vulnerability being discovered to it being weaponised is now well below 24 hours, against an enterprise remediation window of 30 to 60 days for critical external vulnerabilities; nearly 40,000 CVEs were published in the first half of this year alone; phishing was the initial access vector in 23% of investigated intrusions, up from 7% the year before. Microsoft’s summary is that “AI is changing the physics of cybersecurity.” It also insists, in the same document, that “the fundamentals still shape the outcome” — which reads less like reassurance than like a man telling you the brakes work while describing the hill. One more thing about the two models named. They belong to the two companies that signed the White House accord last Tuesday, committing to evaluate their own models for exactly this class of risk. The evaluation happened. This is what it said.
The Market Read the Same Report and Priced It at $2.5 Billion
Twenty-four hours either side of that report, Armadin closed a $255.5 million Series B at a valuation above $2.5 billion, co-led by Andreessen Horowitz and Accel, with Bain Capital Ventures and Redpoint new on the cap table and 8VC, Ballistic Ventures, Google Ventures, In-Q-Tel, Kleiner Perkins and Menlo Ventures all returning. That is roughly $445 million raised in total, about seven months after the company came out of stealth. What Armadin sells is the thing Microsoft just described: specialised AI agents operating as an autonomous offensive swarm, finding vulnerabilities and chaining them together against a customer’s own systems. The founder is Kevin Mandia, who built the incident-response firm that became Mandiant in 2006, ran FireEye after it bought Mandiant for $1 billion, and sold Mandiant to Google for $5.4 billion in 2022 — a man who has spent thirty years arriving after the breach and has now raised nine figures to arrive before it. His pitch is one sentence: “The only way to build a defense that keeps pace is to train it against the best offense available, every day.” The uncomfortable corollary is that somebody has to build the best offense available first, and the business model requires it to keep getting better.
OpenAI Fired Three People From Its Safety Organisation Over What They Did With a Document
The company confirmed on Friday that it “parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” and added that “our investigation confirmed that these individuals mishandled sensitive company information outside established company procedures, violating our policies and breaking the trust essential to our work.” Reporting names them as Jasmine Wang, Tomek Korbak and Mikita Balesni; all three had worked on safety, and all three had previously said in public that AI is moving too fast. Per Bloomberg’s sourcing, the material concerned OpenAI’s own infrastructure architecture, and at least some of it went to an outside organisation that evaluates AI systems — unnamed in every account so far. OpenAI has been explicit that this was not a dismissal for raising safety concerns, and that distinction is doing a great deal of work: the complaint is not what they said, it is which building they said it in. We cannot adjudicate it and will not pretend to. What is on the record is the shape. The accord those same executives signed in Washington four days ago describes internal teams verifying that controls operate as intended. Three of the people who did that work no longer work there, the organisation they allegedly shared with is the kind of organisation the accord’s own logic depends on, and the first week of October is where those two facts meet.
Source: The Hacker News, October 2, 2026
A New Nonprofit Launched to Publish the Part of Model Training Nobody Shows You
Trillium Labs went live on Thursday, founded by Nathan Lambert — who ran the Olmo and Tülu open-model lines at the Allen Institute for AI, and whose congressional testimony on open models we covered here last month — together with Tom Zick, with initial backing from Halcyon Futures and Schmidt Sciences. The specific promise is narrower and better than “open source AI”: it intends to release “fully open post-training recipes, including the data, code, evaluations, and intermediate checkpoints that let other researchers study how model behavior develops.” That last clause is the whole point. Open weights give you a finished model and no way to ask how it got its habits; intermediate checkpoints give you the habits forming. The stated reason is institutional rather than technical — “we believe these problems require a scientific community with the resources to investigate them independently” — and sitting where it does in this week’s news, next to a lab whose internal verifiers are now ex-employees and a threat report assembled from behaviour nobody outside the vendors can reproduce, that is less of an abstraction than it was in September. It is a new organisation with two founders and a blog post. It is also the only story here proposing that anyone outside a company be able to check.
Source: Trillium Labs, October 1, 2026