Top items
- Nathan Lambert revisits his “open models aren’t a cyber catastrophe” claim amid “Black October,” a wave of critical vulnerability disclosures (PHP2Shell, GitLab) and crypto hacks, and largely holds his position while adding nuance.
- Epoch AI demonstrated that the open-weight GLM-5.3 can autonomously turn publicly visible Firefox patches into a working multi-stage browser exploit for ~$1,000 per successful attack.
- OpenAI-backed Thrive Holdings agreed to buy security firm ADT for $15.3B enterprise value, framed as a template for deploying frontier AI across entire organizations.
- AgiBot unveiled the AHA3 humanoid, the GO-3 embodied foundation model, and a new “Worken” robot-utility metric at its Innovation Day.
- AI Safety Feed surfaces new work on lab culture synchronizing insider beliefs and a chain-of-thought reconstruction of a model’s internal charter.
Field & industry developments (cybersecurity and AI)
Nathan Lambert reassesses open-weight models and cyber risk after “Black October.” Lambert’s Interconnects post is a reflection on a take he published a few weeks earlier, written in the context of what he previously coined “Black October” — an onslaught of critical vulnerability disclosures including PHP2Shell and GitLab that, in his view, marked the moment the cyber risks of AI models began to materialize. As he was writing, news broke of another major crypto security incident at Binance, though he notes the crypto industry’s long history of hacks makes it a poor indicator of AI vulnerabilities translating into new societal harm. His read on the behind-the-scenes situation: defensive work has accelerated, with more people and resources being brought in to “patch the internet,” and the Wall Street Journal reported that cyber first responders are strained — which he interprets as “the system working.” He frames the current moment as living through an exponential rise in vulnerabilities that will take a long time to level off at the top of a sigmoid curve.
The core of his argument revisits the fear around Claude Mythos. When Mythos was announced, it was previewed as a new class of cyber weapon that could cause mass societal destabilization if it fell into the wrong hands. Lambert argues this appears to have been wrong: by all measures, GLM-5.3 is the model that actually crossed that capability threshold, its weights have been public for over a month, and there’s little public evidence much changed. He goes further — if Claude Mythos had been accidentally released as open weights, he thinks the world would have been “more or less fine,” producing an acceleration of cyber incidents rather than a step change in risk. He characterizes leading open-weight-fear proponents as having made a falsifiable prediction (that current open models would cause unprecedented AI harms by crippling cyber infrastructure) and argues they are set up to be wrong. To document the alarmism, he cites: Rob Wiblin (80,000 Hours, April 10, 2026) saying Mythos in criminal or hostile-state hands “would be an actual disaster”; Anthropic’s Project Glasswing blog (April 7, 2026) warning the fallout “for economies, public safety, and national security—could be severe”; and Zvi Mowshowitz (April 9, 2026, whom Lambert calls more reasonable) predicting “a cornucopia of zero-day exploits for essentially all the software on Earth … It would be chaos.”
Lambert says he largely stands by his earlier writing but is nervous about it, stressing he’s not a cyber expert. He walks back the word “fine,” acknowledging it undersells a coming “forced industry update to a new way of working with software” that will be expensive and painful; if you’re already somber about AI’s challenges, “fine” is suitable, but if you hadn’t priced in turbulence, it undersells. His actual view is that open models provide balance in the offense-defense equation, and that Mythos arriving sooner wouldn’t have massively upended that balance — he suspects defenders (well-resourced companies) would adopt open-weight models faster than attackers, who are mostly lower-resourced. A key variable he’s watching is how willing adversaries are to cause physical harm in the U.S. and allied countries; so far adversaries have likely used these models and exploits to extract information clandestinely en masse, but it’s unclear whether any party will push the full force of these models against public institutions. He cites a tweet arguing threat actors behave like businesses with defined processes (“a bakery isn’t going to start making dog food because there’s a dog food shortage”), so ClickFix and phishing actors won’t drop their methods to vibecode exploits just because a critical CVE appears — the only actors exploiting these are the ones whose MO already is exploitation. He expects most vulnerabilities are currently being found by closed models (with high error bars), notes open models are expensive to serve and the compute to serve the largest is “extremely sold out” (so a near-term open-model hack would likely come through a poorly secured open-model API), and predicts margins/prices on open models will fall as compute becomes abundant. He emphasizes that this is more about AI’s general capability and diffusion than open-vs-closed, pointing to cybersecurity writers (Joshua Saxe, Chris Hughes) who focus on ecosystem-level trends showing cyber damages growing exponentially since long before ChatGPT. He closes by arguing closed models have played a heavier part in recent hacks than open ones, that cyber is a societal problem AI companies alone can’t solve, and that too much AI-safety effort is going into blaming open models (for which there’s extremely low evidence). Thanks to Seb Krier for feedback. In follow-up Substack notes, Lambert reiterated: “open-weight mythos still seems fine… it could be helping many companies adjust to the new cyber warfare reality,” and that open models will be used for cyberattacks just as closed models have, but they aren’t a “super high leverage tool magically increasing cyber crime by 100x — those bottlenecks lay elsewhere in human factors.”
Epoch AI: an open-weight model weaponized public Firefox security fixes into a working exploit. Corroborating and grounding Lambert’s discussion, Epoch AI published results from internal work using GLM-5.3 to find vulnerabilities and chain them into an attack. In the demonstrated case, they used GLM-5.3 to create a website that, simply when visited by a Firefox user and without any further interaction, gave the attacker the ability to exfiltrate files from the user’s computer. Lambert highlights this as an illustration that real-world harm requires multiple things to align — the exploit is in Firefox, and a would-be attacker still needs to plant a website looking for that exploit; finding the exploit is only one step in a multi-stage process. Per the AI Safety Feed summary of the Epoch post, the model “autonomously turned publicly visible Firefox fixes into reliable, multi-stage browser exploits for roughly $1,000 per successful attack.” The broader warning: as cyber-capable models improve, published patches may rapidly become attack blueprints unless developers substantially shorten or conceal the patch-deployment window (i.e., the gap between a fix being published and reaching users becomes a critical exposure window).
Company & product developments
OpenAI-backed Thrive Holdings to acquire ADT for $15.3B to retrofit it with AI. On October 29, Josh Kushner’s Thrive Holdings struck a definitive deal to buy home-security company ADT at $10/share — roughly $7.6B in equity, $15.3B in enterprise value, representing a 45.6% premium, with a 30-day go-shop period. OpenAI is Thrive’s partner and investor, supplying frontier models and forward-deployed engineers for the integration. OpenAI COO Brad Lightcap framed the deal as a template for “frontier AI deployed rapidly across entire organizations.” Apollo is financing the transaction with no financing contingency. (Source: TechCrunch via AI Weekly.)
AgiBot Innovation Day: AHA3 humanoid, GO-3 foundation model, and a new “Worken” metric. AgiBot bundled a stack of launches at its Innovation Day (Oct 29): the AHA3 humanoid platform featuring liquid cooling and IP67-rated joints; the GO-3 Preview embodied foundation model (built on a “WRAM” framework), which the company says ranks #1 on the RoboDojo benchmark and highest on RoboTwin; the omni-modal WITA-Omni 1.0 interaction model; the G2 Max heavy-duty dual-arm industrial robot (21 degrees of freedom, 40kg payload per arm); and Link-U OS 1.0, which AgiBot plans to open-source by the end of 2026. Alongside the hardware, AgiBot introduced “Worken,” a proposed utility metric defined as active-count × autonomous-hours × output-efficiency, meant to replace unit shipments as the real-world measure of robot usefulness. (Source: roboticsandautomationnews.com via AI Weekly.)
Research & AI safety
Lab culture can synchronize insiders’ beliefs and priorities (Less Wrong, Oct 30). A forum post argues that while AI-lab employees gain valuable frontier information, they may also absorb organizational incentives that subtly synchronize their attention and beliefs. The author contends the AI safety field needs more independent observers, and recommends that insiders take breaks, listen to outsiders, and sometimes leave labs to preserve epistemic independence. (Tagged: epistemics/rationality, alignment field-building, Anthropic.)
Chain-of-thought reconstruction of a model’s internal charter (Less Wrong, Oct 29). A researcher reconstructed much of “Meta Muse Spark Open’s” apparent internal Charter by eliciting repeated passages from the model’s chain of thought. The reconstructed specification reveals a corrigibility-first hierarchy, fiduciary duties to users, explicit evaluation-awareness guidance, and a distinctive approach to AI discretion and representation — a demonstration that a model’s governing spec can be partially extracted through its reasoning traces.
Toby Ord on AGI timelines and recursive self-improvement (80,000 Hours Podcast). Ord argues that recursive self-improvement could create dangerous acceleration but is neither inevitable nor necessarily an instantaneous intelligence explosion. He advocates preserving human oversight, establishing targeted safety norms and moratoria, pursuing verifiable international coordination, and planning across wide uncertainty about AI timelines.
Spencer Greenberg on sustainable mental health in existential-risk work (80,000 Hours Podcast). The discussion notes that people working on existential risks often experience chronic stress, guilt, burnout, and isolation even absent clinical anxiety or depression, and argues that rest, boundaries, holding multiple values, and supportive relationships should be treated as performance-enabling necessities rather than moral failures.
Commentary & analysis
Nathan Lambert on org structure as a hidden determinant of model quality. In a Substack note, Lambert argued it is “vastly underrated how important org chart and team functionality is to building a good LLM,” pointing back to a favorite Interconnects post over 18 months old. He says this is why organizations crumble and need rebuilding (citing Llama/xAI as examples) and why underdogs can still catch up: an effective org and team is “worth multiple multiples in compute.” He added that there isn’t much secret sauce to building models — beyond a few tricks, most gains come from consistent execution and using compute wisely.
AI in education: Alpha School and the limits of EdTech (Zvi Mowshowitz). Within a broader essay on school dysfunction, Zvi examines Alpha School’s AI-driven model, drawing on Kelsey Piper’s reporting. Alpha students do roughly two hours of academic drill a day using the “Timeback” EdTech package and appear to outperform kids spending far more time. Piper argues Alpha’s advertised “learning multiplier” numbers are bogus and its MAP-score gains are inflated, but concedes no one took the multipliers literally; the real “magic,” she says, is the environment and incentive system that reliably gets kids to use the software — a low student-teacher ratio, teachers present mainly to keep kids on track, and prizes (even expensive ones like a Nintendo Switch or iPad, which Zvi notes costs under 1% of a semester) earned for finishing work. Piper was “broadly unimpressed” with the publicly available apps: she criticizes Alpha’s “baffling and ill-advised decision” to use AI-generated text for most reading-comprehension and content-knowledge instruction — citing TeachTales, which churns out bite-sized Harry Potter fanfiction gated behind comprehension questions, which she found “deeply depressing.” Zvi partly defends the approach, hypothesizing that AI-generated passages can be customized to each child’s exact comprehension gaps, grammatical structures, and story preferences, and can fold in history/science content in ways real books can’t be adjusted on the fly, while agreeing it’s no substitute for reading actual books. When others deploy the Timeback software without Alpha’s environment, results are weaker — but Zvi argues Piper undersells those outcomes (underperforming groups improved, just not spectacularly). His overall diagnosis: Alpha isn’t genius so much as “the one-eyed man in the kingdom of the blind” — nearly the only school not wasting most of the day — and its high price (up to ~$75,000) looks less extreme next to the ~$36k average (and $52,349 at PS 307) that other schools spend for worse results. (The surrounding essay’s extensive K-12 policy material — PS 307, teacher unions, funding theory, grade inflation — is largely outside the AI scope of this digest.)