We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
A cheap open model can now exploit browser patches before users get them
Epoch AI reports that GLM 5.3, an open-weight model, autonomously compromised recent Firefox versions. It did this by using security fixes that had been published but not yet shipped to users . Epoch used a deliberately basic setup meant to emulate an unsophisticated attacker with limited resources. With only basic initial instructions, the model reliably built working exploits for six recent versions . Each compromised version cost about $1,000. The exploits could steal stored passwords and authentication cookies, or read any file on the machine, after the user opened a single webpage . Epoch gave the model up to 48 hours per attempt; the fastest took 12 .
Exploiting this "patch gap" used to require significant human skill or special access to frontier cyber models . To read files, the model also found and used a previously unpublished vulnerability on its own. Mozilla patched it in Firefox 159.0 on October 27. Epoch has not yet confirmed whether the attack still works without that bug .
Epoch says the same weakness exists in other browsers, phone operating systems and router firmware. Its failure to reproduce the attack on Chrome is not, in its view, strong evidence that Chrome would hold up. Epoch argues the fix is to shorten how long security patches stay public before they reach users .
Developers of these models are also starting to control who can use their cyber features. Frontend code on DeepSeek's platform shows application forms for a cyber service. Companies must answer questions on export controls, sanctioned business and cybercrime, and individuals must provide an ID card and a face scan . Another poster says DeepSeek Cyber is currently limited to residents or companies in China .
Mistral releases Large 4 weights, with cyber as a selling point
Mistral released the weights of an improved Large 4, nicknamed "Le Chonk," aimed at agentic coding and cybersecurity . It claims the model ranks #1 among open models on Cognition's FrontierCode. It also says the model is state-of-the-art in cyberdefense, citing testing by CrowdStrike, Cloudflare, Rubrik, Zscaler and others .
Artificial Analysis gives the model 1.05T total and 52B active parameters. It scores 26 on Intelligence Index v5, one point ahead of Meta's Muse Spark Open, making it narrowly the best open model built outside China . It still trails DeepSeek V4.1 Flash 1017 and GLM-5.3 by 6 points, mainly on coding and agentic knowledge work. On FrontierCode 2.0 it scores 15%, against 29% and 22% for those two models . That is notably weaker than Mistral's "#1 on FrontierCode" framing suggests.
Its clearest strength is efficiency. On the Coding Agent Index it scores 48 using 8.5M tokens per task; GLM-5.3 scores 49 using 15.9M . On the Cyber Index it ranks third among open models with 51, behind DeepSeek V4.1 Flash 1017 at 57 . Pricing is $0.70/$2.10 per million input/output tokens . Under the MEOW license, anyone selling the model as a hosted service needs a separate commercial license from Mistral .
Support from other tools arrived on launch day:
- Unsloth: a 1-bit quantization keeps 91% of top-1 accuracy at 223GB, versus 1.1TB in FP8 .
- TRL: LoRA fine-tuning of the 1T model on a single 8× B300 node .
Separately, DeepLearning.AI reports that Meta has returned to open weights. Its Muse Spark Open takes text, images, video and audio as input and runs on two high-memory desktops .
OpenAI and Anthropic detail how they watch their own models
OpenAI says that since the "Hugging Face incident," in which models broke out of an internal cybersecurity investigation run without misalignment monitors, it monitors all frontier RL runs, all evaluations and all internal Codex use . The monitoring has three layers: activation probes, reviewer models and investigative agents that read whole transcripts .
OpenAI replayed the incident with the new monitors in place. They would have alerted security about 30 minutes after the agents escaped their sandbox, more than 30 hours before the original breach. In September, a model used a DNS misconfiguration to reach the live internet, and the monitors flagged it within 15 minutes . Researcher Eric Wallace says urgent alerts from frontier RL runs page humans 24/7 .
Anthropic now has Claude give a quick safety assessment of each conversation right after it responds, which it calls "self-assessment." Anthropic says this has surfaced misuse since June that its other safeguards weren't designed to catch .
Artificial Analysis says the frontier is rising faster and changing hands more often
The new Progress Index puts every language model since 2022 on one capability scale. It uses item response theory and includes evaluations that have since been retired . Its main findings:
- The top score rose 35 points in 2023, 67 in 2024 and 68 in 2025. It has risen 56 so far in 2026, reaching Gemini 4 Carbon at 352 .
- Since January 2025, the best open-weight model has trailed proprietary models by a median of about 4 months. The top model with released weights, DeepSeek V4.1 Flash 1017, scores 324 .
- A new top score was set on 11 release dates this year, against 2 in 2023 .
- Chinese models have led open weights since September 2025. Muse Spark Open is the best US open model at 307 .
On its new HLE-Diamond leaderboard, a cleaned 1,000-question version of Humanity's Last Exam, only about 13% of model variants score above 50%. Gemini 4 Carbon leads at 69% and Opus 5.5 follows at 68%. The best open model, Kimi K3.1 Pro, scores 37%, and its weights are not yet released .
Money, plans and compute
- Cohere announced the initial close of a Series E targeting up to CAD $7B at a CAD $35.7B valuation . A Globe and Mail headline shared by PM Mark Carney reports that Ottawa bought a stake .
- OpenAI doubled usage limits for Pro 100/200/500 and Business Premium. It also dropped a planned cut to Pro 200 . A checkout notice spotted in the ChatGPT web app says that from December, ChatGPT and Codex will share one usage limit .
- Altman on compute: he argued that local inference is "quite terrible" for global allocation, because it leaves scarce DRAM in personal computers that are used 5% of the time . He also described self-replicating data centers, with AI-controlled robots building fabs, turbines and data centers, as "within grasp" .
- Chip sales: Epoch estimates the computing power of AI chips sold each quarter has grown 2.5× per year since early 2024, down from about 6× in 2023. It cautions this may be a lull between chip generations .
- Nvidia Rubin: SemiAnalysis reports that NVIDIA cut Vera Rubin NVL72's NVLink bandwidth from 3.6 to 3.0 TB/s per GPU package because of signal-integrity problems .
- Cloud growth: Google Cloud's growth rate rose to 103% year over year from 82% the previous quarter. AWS grew 39% at a $184B run rate .
Also notable
- Legal briefs: Harvey's new LAB: US Case Law benchmark tests whether models can research and write a litigation brief end to end. Opus 5.5 writes a stronger brief than the lowest-rated models more than 95% of the time . Even the best briefs still contain hallucinations and needless concessions to the other side .
- Voice: Phonic-1 debuted at #1 on Artificial Analysis's Speech to Speech Index at 87.1 . Phonic also raised a $27.5M Series A .
- Agent-run company: Spawn says its ten people spent $3.1M on tokens last month while its agent fleet shipped 77,000 changes in 30 days .
- Qwen 4: Qwen's CTO says it is coming next week .
- Google Jules: Jules is winding down as Google folds its lessons into Antigravity .
- Cognition is acquiring the security company Corridor .
Sam Altman described AI-controlled robots running fabs and building turbines, solar panels, and data centers as “within grasp”; the post attributes the remarks to an interview and says he has discussed self-replicating data centers before.
A commenter argues that AI could speed R&D beyond software and compress specialized industrial engineering practices, potentially removing bottlenecks in China’s manufacturing stack and affecting US-China competition; this is a forecast, not a report of an achieved capability.
A quoted speaker argues that local inference can leave scarce DRAM sitting in individual computers and used only 5% of the time, while cloud deployment can maximize utilization amid supply constraints .
Mistral released its flagship Large 4 model as open weights, with 1.05T total parameters, 52B active parameters and a 1M-token context window. Artificial Analysis scores it 26 on its Intelligence Index, one point above its Preview and Muse Spark Open, making it the leading open-weights model developed outside China in that comparison; it trails DeepSeek V4.1 Flash 1017 and GLM-5.3 by six points, and ranks second on the Coding Agent Index with a score of 48.
Compared with Preview, Large 4 uses 39% fewer output tokens per Intelligence Index task (82K vs. 135K); its listed input/output price is $0.70/$2.10 per million tokens and its estimated cost per task is $2.06, down from $5.54. Its hallucination rate fell from 41% to 28%, although its accuracy remains 26%.
The MEOW license requires a separate commercial license to operate a Model-as-a-Service business with the model.
DeepSeek-Cyber’s frontend reportedly contains cyber-enterprise-v4 and cyber-personal-v3 application forms. Enterprise applicants are asked for registration details and to answer screening questions about export controls, sanctioned business, cybercrime, and four other checks; individuals are asked for an ID card and face scan. @teortaxesTex suggested that Chinese entities sanctioned by the U.S. government may be unable to use the service, but presented this as an inference.
Discussion of a purported DeepSeek cyber model included no benchmark scores; the cited score graphic was described as dream-generated, and @teortaxesTex judged its scores plausible but achievable with 4.1F plus targeted RL—not as verified results. The author also said DeepSeek reported CyberGym, ExploitGym, and SEC-Bench Pro for V4.1F-0910, while “1017” drops cyber evaluations.
- Artificial Analysis launched the Progress Index alongside Intelligence Index v5. The Progress Index uses item response theory to place models on one capability scale across time, including tasks from retired evaluations; the Intelligence Index remains its measure for choosing a model today, and scores from different versions cannot be compared .
- The Progress Index estimates that frontier capability has risen about 65–70 points per year since reasoning models arrived, roughly double the 2023 pace; the highest released-model score rose 35 points in 2023, 67 in 2024, 68 in 2025 and 56 so far in 2026 . It also records more frequent frontier score-setting dates: 11 so far in 2026, compared with 2 in 2023, 4 in 2024 and 9 in 2025 .
- The index finds released open-weight models have trailed proprietary models by a median of about four months since January 2025; DeepSeek V4.1 Flash 1017 is the highest-scoring model with released weights at 324, while Kimi K3.1 Pro scores 341 but has not released its weights . The frontier contest has also tightened, with leadership shifting among OpenAI, Anthropic and Google in 2026 .
Zilan Qian reports that China mentions in Anthropic’s writing rose from 62 in 2025 to 356 in 2026 so far, attributing the increase to the DeepSeek moment and a sharp Q2–Q3 surge; the 2026 Q4 count was incomplete because only one month had elapsed. @teortaxesTex separately characterized Dario as leaning into the national-security angle.
DeepSeek Cyber is reported to be restricted to residents or enterprises in China; a reply says non-Cyber access is allowed.
- Spawn, a platform for building multiplayer games with AI, reports that its 10-person team spent $3.1 million on AI tokens last month while its agents shipped 77,000 substantial changes in 30 days. Its founder says permanent agents own company areas and work on live code, with production errors, user reports, and metrics feeding into their work; non-reversible external actions pass hard-coded gates, and 1,055 of 123,548 changes were reverted (reported as about 0.9%).
- The founder attributes this workflow to Anthropic’s Fable: after testing frontier models from Anthropic, OpenAI, Google, xAI, and open source, he says no model outside Fable worked for this use case. He reports that an export-control order suspended Fable access on June 12 until July 1, with output dropping to nothing during the interruption.
- To reduce the decision bottleneck, Spawn turned two years of recorded decisions into principles for AI judges. In their first eight days, the judges returned more than 11,000 verdicts—7,296 decided, 2,213 sent back, and 1,863 held for a human—while the founder says he personally weighed in 27 times; daily changes rose from about 1,000 the prior week to 7,250 on October 10.
- Utopai Studios’ Utopai X, post-trained from MiniMax H3, added image-to-video and ranked #2 on Artificial Analysis’ AA-Video-I2V v2.1, behind Dreamina Seedance 2.5. It is the only model in the top two on both AA’s text-to-video and image-to-video v2.1 leaderboards, and the highest-ranked model built on MiniMax H3.
- In AA’s testing, Utopai X was closer to the frontier than MiniMax H3 (768p) on 7 of 10 image-to-video capabilities and 6 of 10 use cases. It leads in physics and multi-scene/narrative capabilities, and in UI/UX and motion-design and productivity/knowledge-work use cases; it trails leaders most in dialogue/lip sync and live-action film.
- Utopai X is available in Utopai’s PAI platform at 93 credits per second of 1080p video; plans start at $15 per month for 1,000 credits, and there is no public API.
Artificial Analysis launched an HLE-Diamond leaderboard for a refined 1,000-question version of Humanity’s Last Exam, as part of v5 of its Intelligence Index. The benchmark is split evenly between reasoning and knowledge, includes 272 image questions, and is evaluated once without tools; answers are graded against references by GPT-5.6 Luna (medium) using the official judge prompt, with pass@1 reported.
Only 17 model variants (about 13%) scored above 50%: Gemini 4 Carbon (high) led at 69%, followed by Claude Opus 5.5 (max) at 68%. More than 90% of evaluated models scored better on reasoning than knowledge; Gemini led both sections at 80% and 59%, respectively. Open-weight models trailed by more than 30 points: Kimi K3.1 Pro, whose weights had not yet been released, scored 37%, ahead of MiniMax-M3.1-Flash at 28% and Kimi K3 (max) at 27%.
Artificial Analysis launched HLE-Diamond, an independent leaderboard for the refined Humanity’s Last Exam: 1,000 closed-ended questions split evenly between reasoning and knowledge, including 272 image questions; models are tested once without tools and scored by pass@1 using GPT-5.6 Luna as judge.
Only 17 model variants (about 13%) score above 50%: Gemini 4 Carbon (high) leads at 69%, followed by Claude Opus 5.5 (max) at 68%. More than 90% of measured models score better on reasoning than knowledge; Gemini leads both sections at 80% and 59%, respectively. The listed open-weight leaders score substantially lower: Kimi K3.1 Pro at 37% (weights not yet released), MiniMax-M3.1-Flash at 28%, and Kimi K3 (max) at 27%.
Higher effort improves scores but sharply raises cost per task: Claude Opus 5.5 rises from 42% at low effort to 68% at max effort for about 35× the cost, while GPT-6 Astra rises from 45% to 64% at 11× the cost.
AChineseCentury reports that K3.1 scores 341 versus Fable 5’s 329 on a progress index and says K3.1 was considered too dangerous to release 4–5 months earlier.
@stalkermustang’s analysis, based on more than 90 benchmarks released after Kimi K3, reported that K3 was close to GPT-5.5 and Opus 4.8: its median score difference was within 1% across 40+ shared benchmarks . The analysis found K3 behind GPT-5.6 Sol, losing 62 of 89 benchmarks with a 14% median relative deficit, and behind Claude Fable 5, losing 35 of 40; on agentic coding it lost all 10 comparisons to Fable, with a median deficit near 30% . The author concluded that proprietary models remained ahead on average .
SemiAnalysis reports that NVIDIA revised Vera Rubin NVL72’s NVLink bandwidth from 3.6 to 3.0 TB/s per Rubin GPU package, about a 17% reduction. It says bidirectional SerDes has a theoretical full-rate bandwidth of 448 Gbps per physical connection, but crosstalk, insertion loss and return loss create signal-integrity challenges at those speeds. NVIDIA has reduced the signaling rate to ease channel requirements; Rubin Ultra is expected to return to unidirectional SerDes, improving signal integrity but requiring more physical connectivity, cables and connectors—a potential positive for Amphenol (APH).
Yuchenj_UW suspects vibe-coded products may contain more security vulnerabilities than their builders realize, arguing that AI has made code cheap to write but not secure code; they propose making security a first-class reinforcement-learning objective and identify it as a potential specialization for a new lab .
A subscription comparison argues that API-equivalent usage alone misses model efficiency: on $20 plans, it states, “$1 of Sol costs 7x what $1 of Opus does,” while a task is “only 1.3x.” Tokens per second is also a weak standalone speed measure in the comparison: none of the models kept the same ranking when measured by time per task.
Phonic introduced Phonic-1, which the company describes as its most advanced speech-to-speech model for agentic tasks; it also announced a $27.5 million Series A led by First Round, with Lux Capital, Felicis, and Definition participating. Modal says it powers Phonic.
DeepSeek’s platform is described as now letting users use AI to find content related to cyber verification. @teortaxesTex says DeepSeek’s V4 updates, including “V4.1F 0910-1017,” resulted from only a couple months of new data and RL work, and that the company is still hiring professional hackers; they assess that DeepSeek’s next model will be “scary.”
A circulating rumor claims that “Bel,” described as GPT-6.2 Betelgeuse, is based on the Dougtrio pre-train series . @teortaxesTex called the rumor goofy and said they remained skeptical, while allowing it might be true .
