← Dailies

Daily · 2026-10-31 00:00 UTC

An AI agent operated on a mouse brain; GLM 5.3 turned patches into exploits

An agent picked a gene therapy, a brain target and the injection coordinates, then performed the surgery — and refused to continue when it detected a calibration fault. Plus GLM 5.3 turning Firefox patches into working exploits, and a shared robots.txt opt-out that the big AI labs just agreed on.

Source ↗

1. An AI agent chose a gene therapy and operated on a mouse brain unaided — A researcher at Oxford gave Claude Sonnet 5 nineteen candidate gene therapies, an epilepsy mouse model, and a motorized surgical device with cameras, with one instruction: reduce the seizures. It selected the drug, the brain region and the stereotactic coordinates itself, deriving injection positions by averaging three papers, and made 32 observations of the surgical field during the procedure. A licensed veterinarian approved every irreversible step and held an emergency stop. When a dropped connection put its calculated target outside the device's physical range of motion, the agent refused to continue — and was right, because manual adjustments during a pause had thrown off the control system. Opus 5.5 and Opus 5 declined to attempt the procedure at all. Results are not peer-reviewed. The Decoder · Oct 30
2. GLM 5.3 built working Firefox exploits from the patch gap — Epoch AI researchers found the open-weight model exploited the window in which a security fix is visible in code but has not reached users, across six Firefox versions released between September 17 and October 26, at roughly $1,000 in API fees per successful attack. The fastest chain took twelve hours against a 48-hour cap. Neither Z.ai nor other providers triggered safety guardrails or refusals. Mozilla patched the underlying vulnerability in Firefox 159.0 on October 27. The same setup against Chrome burned about $100,000 without producing an equivalent exploit. The Decoder · Oct 30
3. OpenAI, Google, Microsoft and Anthropic back a shared scraping opt-out — The four labs say they will adopt a "policy token" signal, proposed by Anthropic's Kevin Kelley at the IETF, that lives in a site's robots.txt so publishers can state AI training preferences once instead of company by company. Anthropic has adopted it already; Google and Microsoft plan support next year, and OpenAI says the coming months. Meta's communications head Andy Stone signalled support but did not say whether the company will honour the token. The Verge · Oct 30
4. Apple Intelligence arrives in China with iOS 27.3 — An iOS 27.3 beta on November 4th brings notification summaries, real-time translation, rewriting and proofreading tools, visual intelligence and object removal from photos, with a full release in December. The new Siri is not part of this launch. Apple says it worked with Alibaba to add support for its Qwen chatbot in some experiences. Apple Intelligence was approved in China in July. The Verge · Oct 30
5. Google's first data-center satellite is online — The Suncatcher prototype carries four Tensor Processing Units and generates up to 1 kW, but the power draw means it can only run one chip for about twenty minutes at a time. Its first task: writing a haiku about being in space, and reporting on orbital conditions around the spacecraft. The test campaign will look at thermal stress and how radiation affects the chips in reality rather than simulation. Two more satellites with optical inter-satellite links are planned for 2027. TechCrunch · Oct 30
6. Anthropic expected to get a high junk credit rating after its IPO — The Financial Times reports Blackstone, anchor investor in a recent $63 billion deal to finance Anthropic's chips, told investors it expects a double-B rating — the top tier of speculative grade — reversing its earlier forecast of an investment-grade rating before 2027. Both Anthropic and Blackstone declined to comment. A junk rating would make it costlier for the lab to raise money against its compute commitments. TechCrunch · Oct 30
7. Samsung warns the AI memory shortage will worsen through 2027 — Samsung said demand for memory will "far exceed" supply in 2027, with AI infrastructure capital expenditure reaching what it called a historic level. Memory sales rose 16 per cent quarter over quarter and 427 per cent year over year, and the memory business posted record quarterly revenue and operating profit. Supply growth kept getting outpaced by demand, the company said on its earnings call. PYMNTS · Oct 30
8. The AI sell-off took nearly $1tn off the Nasdaq in a week — Friday's session drove US markets down by hundreds of billions, leaving the tech-heavy Nasdaq nearly $1tn lower on the week and around $2tn lower on the month — an October wipeout slightly larger than SpaceX's entire market value. Intel fell almost 30 per cent on guidance, flirting with its worst day since 1974, and Nvidia dropped 3.1 per cent. A day earlier Alphabet lost more than 12 per cent, wiping $500bn, after reporting heavier cash burn on AI investment. The Guardian · Oct 30
9. Sebastian Thrun raises $130M for hardware-design AI startup Dulo — Dulo, co-founded by the Google X co-founder and Adrian Aoun, has raised $130 million from backers including former Google chief executive Eric Schmidt, early Google investor Ram Shriram, Salesforce's Marc Benioff and venture capitalist John Doerr. The company's one-page site says it is building foundation models optimised for hardware design tasks, aiming at "manufacturing at lightspeed." It did not disclose a valuation. Bezos-backed Prometheus is the best-funded rival in engineering AI. SiliconANGLE · Oct 30
10. Krafton's Ludo Robotics ships a robot brain built around human intent — Ludi 0.2, from the Krafton subsidiary headquartered in Palo Alto with an R&D lab in Seoul, adds social intelligence, memory, reasoning, tool coordination, gesture interpretation and physical movement to the vision-language model in Ludi 0.1, released in August. The stated goal is a single "brain" that watches, listens, interprets gestures and reasons about an environment in real time — the company's route from game interaction to robots in retail and domestic spaces. SiliconANGLE · Oct 30

Try This Next

• Probe the patch gap in your own stack: pick a dependency with a public security fix and ask a model what it can infer from the diff alone.
• Check whether your site's robots.txt expresses AI training preferences once, rather than relying on per-crawler rules.
• Save Sonnet 5's mid-procedure refusal as a case study — a capable agent declining an unsafe action is a feature to design for, not a bug to route around.

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.