AI digest — Nov 2, 2026
Top line: MIT researchers report using GPT-6 Astra to edit Qwen's open weights directly — weight-space programming lifting the small model's MATH-500 score from 7.9% to 48.1%; publishers sue Brave over AI scraping, and OpenAI names Dan Driscoll as senior national-security counselor while more than quadrupling critical bug-bounty rewards.
Brief
- MIT report: GPT-6 Astra edits Qwen's open weights. VentureBeat: MIT researchers used OpenAI's GPT-6 Astra to directly edit the open weights of Qwen3.5-0.8B — a technique the piece calls weight-space programming, "like gene editing for neural networks" — raising the model's MATH-500 score from 7.9% to 48.1% while touching roughly 0.016% of the weights, with no retraining. The single earlier claim of this experiment circulated only as a relayed X post; this is the first proper report, but the underlying paper and MIT's own confirmation aren't yet in hand.
- Times and other publishers sue Brave over AI scraping. The Verge, as reported by Reuters: The New York Times plus the companies behind the New York Daily News, the Chicago Tribune and the Orlando Sentinel have filed lawsuits against Brave Software over alleged unlawful scraping used to train AI and for letting users bypass paywalls. Their complaint says Brave has monetized its index of scraped websites in part by selling AI companies paid access to "real-time online content." A new defendant in the publishers' scraping fight; the complaints' full claims and Brave's response aren't yet in hand.
- Anthropic restricts Claude Code game clones. The Verge, as reported by Tom's Hardware: Anthropic has updated Claude Code's system prompt to stop the coding tool being used to recreate existing video games — retro gaming is currently crowded with AI-made decompilations, recompilations and ports — while still allowing people to build games with it. What this does to the clone pipeline, and whether other labs follow, is the open question.
- Study: AI may improve patient discharge summaries. Medical Xpress: a study concludes AI could help physicians prepare hospital discharge summaries more efficiently and with greater accuracy and completeness. The outlet's summary doesn't name the institution or where the study appeared, and no independent read is available yet.
- Google reports $100K-a-day LLMjacking. SiliconANGLE, on a report from Google's Mandiant and Google Threat Intelligence Group: attackers who hijack cloud accounts are running up as much as $100,000 a day in fraudulent charges by using victims' compute to run large-language-model workloads — attackers stealing access to compute rather than the classic haul of data. The figures come from Google's own security teams with one outlet carrying the report so far; the report itself isn't independently checked.
- Music industry backs the Times against OpenAI. Music Business Worldwide reports the music side has entered the New York Times' copyright case against OpenAI: RIAA, NMPA, ASCAP, BMI and five other music organizations filed a proposed amicus brief of their own, formally supporting the newspaper, alongside Irving Azoff's Music Artists Coalition, Cate Blanchett and CAA. The filing follows the Wikimedia Foundation's own bid to weigh in as amicus; the court's response to the pile-up of outside briefs is the open question.
- OpenAI's LifeVal benchmark. OpenAI introduces LifeVal, a benchmark measuring how well AI helps people complete everyday tasks — from finding childcare to managing personal finances and government paperwork — shaped by feedback from thousands of ChatGPT users, and says it has used it internally over the past year to measure and improve its models (OpenAI on X). The company's own announcement so far; how the benchmark is scored, and any outside replication, isn't yet available.
- OpenAI's Grid Protectors Program gets its first pilot site. The Verge and OpenAI: OpenAI will pilot the program at an upcoming data center in Effingham County, Georgia, where it promises to free up to 1,000 megawatts during extreme weather or emergencies — by its own estimate enough to run air conditioners for around 700,000 homes. The announcement lands amid spreading US opposition to data centers. Company claims only so far: no named utility partners, and no independent detail on how the power commitment would be enforced.
- OpenAI hires Dan Driscoll for national security. OpenAI appoints Dan Driscoll as Senior Counselor for National Security Modernization, to advise on responsible AI for national security modernization, cyber resilience, and collaboration with US allies. The company's own announcement so far; no independent reporting on the appointment or his remit is available yet.
- robots.txt training opt-out reported live. Tech Times reports Google, OpenAI, Microsoft and Anthropic now honor a single two-line robots.txt token that blocks AI training on a site across all four platforms at once — the publisher-side opt-out standard the four labs committed to just days earlier, moving from announced commitment to a signal publishers can actually set. The piece stops short of documenting the token's exact syntax or how each platform enforces it.
- OpenAI plans next security sprint and bigger bug bounties. OpenAI (announcement on X) says it is planning its next internal security sprint with outside partner Cathedral Systems, whose researchers will use its models alongside OpenAI's teams to probe its defenses, and is raising bug-bounty rewards — more than quadrupling awards for critical findings, with exceptional findings eligible for up to $172,000 including bonuses. Company announcement only so far.
- Indian Express runs a "what is" explainer on VOLTA-AUTODOXER, walking general readers through the AI-enhanced malware pattern — infostealers that had previously stolen login credentials for major breaches, now integrating an AI stage that profiles and ranks victims. The piece introduces no technical detail beyond what Cisco Talos's analysis already established: which model the ranking stage uses, how widely the malware has spread, and whether other kits copy the design all remain open.
- Tech Times describes HarnessPro, Tsinghua University research that lets AI coding agents rewrite their own scaffolding — reporting a 26% cost cut and accuracy gains with no model retraining, with Factory Droid 1.0 also touched on. The figures rest on one outlet's summary of the study so far; the paper or project page isn't independently checked yet.
- Instinct's WhatsApp disruption. Indian Express reports the personal-AI-agent service says it traced the WhatsApp disruption to a configuration issue, while separate user ban reports have raised questions about what the agent does with accounts. Beyond the company's own account, what triggered the ban reports and how many users were affected remain unconfirmed.
- Goldman on post-midterm AI rules. Seeking Alpha summarizes the Goldman Sachs read from economists Alec Phillips and Abhay Duggirala: AI regulation is expected to advance whatever the midterm outcome delivers, but sweeping new federal restrictions are unlikely in either scenario. The takeaway is analyst-layer anticipation, not a checkable regulatory move.
- Tech Times adds to the reported prototype Meta built with the Shanghai robotics firm Fourier Intelligence: it says the robot was assembled in roughly three months and runs on Nvidia's Jetson Thor computing platform, links the robot to Meta's Muse AI line, and frames the Chinese partner as bound by Chinese intelligence laws — a national-security angle on the tie-up. Neither company has confirmed the prototype.
- Google DeepMind's blog describes work with the UK government aimed at cutting the time planning assessments take. The company's own account so far — no independent reporting and no detail on which assessments, tools or pilot areas are involved.
- Perplexity's open-weight embedding models. Arav Srinivas on X announces pplx-embed-v2-context, open-weight vision-language embedding models in 9B and 0.6B sizes, each returning a single int8 vector per document or per chunk, with a paper co-written with turbopuffer. The company's own post so far — no independent coverage or benchmark yet.
- LeVJEPA 1.5 Large announced. Yann LeCun amplifies a result from his new research group's orbit (LeCun's RTs, second): researcher Lukas Kuhn announces LeVJEPA 1.5 Large, a model that reportedly closes the gap to Meta's V-JEPA on both image and video benchmarks while training with 5.6× less compute and an order of magnitude less data — with LeCun framing it as evidence that improving pretraining losses beats chasing more data and GPUs. Researcher-channel announcement; no paper link or independent benchmark yet.
- OpenAI's Git at scale. OpenAI describes Pando, an internally built Git-compatible backend designed for the scale of development its Codex coding agent makes possible. An engineering post about the company's own tooling; no independent coverage or outside benchmark yet.
- Grok @Bot feature rundown. Elon Musk on X points to Matt Palmer's walkthrough of Grok @Bot improvements, which itemizes shopping cards, product pages, order tracking, payments, custom modes, team-bot logistics and a Slack connector. A feature-drip rundown with no changelog or independent verification yet.
- Qodo brings context to Codex code review. OpenAI says Qodo's plugin pulls company knowledge and coding standards into Codex to catch high-priority issues before a pull request opens. A single integration announcement.