AI News - Saturday, 21 November 2026 - Riemann Proof Claim, Signal AI Block, Secret Leaks
In a Nutshell
An Anthropic mathematician says Claude helped him prove the Riemann hypothesis, while METR's updated data shows frontier capability still doubling roughly every four months. The same week brought the bill: Signal blocks AI screen-reading, two million planted secrets leaked through AI routers, and Meta rethinks its climate pledge as AI data centers strain the grid.
5-minute AI news update - 21 November 2026
In this AI News
Anthropic mathematician says Claude helped him prove the Riemann hypothesis
[Research] Levent Alpöge published a Lean-compilable proof of the 1859 conjecture and says the method extends to the Generalized Riemann Hypothesis. Mathematicians have confirmed the Lean code compiles, but no independent verification of the argument exists yet. Anthropic itself has not endorsed the claim. Source: The Verge
METR's updated time horizon plot shows capability gains holding but judgment lagging
[Research] The evaluation nonprofit added longer and more open-ended tasks to the chart that became the reference point for AI progress claims. Time horizons still double about every four months on verifiable work, while tasks needing judgment improve far more slowly. That gap decides whether agents can automate research rather than just coding. Source: MIT Technology Review
Signal blocks screenshots to stop AI agents reading your messages
[Policy] Signal restricts screenshots on iOS 27 and Android, which stops Siri, Gemini and other assistants from reading on-screen content. President Meredith Whittaker said agents are hostile to privacy and that Apple, Google and Microsoft give developers no practical way to limit data access. She called the remedy incomplete, since agents can still reach app data through the file system. Source: The Verge
Two million planted secrets leaked through AI routing middlemen, researcher finds
[Research] Dragos chief research officer HD Moore seeded unique credentials into single prompts sent through AI routers such as OpenRouter and Vercel. The secrets surfaced across providers, exposing prompt theft and logging failures at each hop in the chain. Any institution sending sensitive prompts through a reseller inherits that exposure. Source: Ars Technica
Meta re-evaluates its 2030 net-zero pledge as AI data center emissions climb
[Industry] Meta is re-evaluating the 2030 sustainability goals it set before the current build-out, with record emissions attributed to AI data centers. Every hyperscaler now faces the same arithmetic between training capacity and carbon commitments. Campus and procurement sustainability reporting will feel this first. Source: Ars Technica
Nvidia board ousted Huang over a year of lying, not the spy allegation
[Industry] Wired reports directors could not establish whether Alice Huang was a foreign intelligence asset, but concluded the former CEO had lied to them across a year. Chair Harvey Jones is using governance grounds to shut down talk of a return. The chipmaker is running its largest capital cycle under an interim CEO. Source: Wired
OpenAI cuts protected reasoning extraction from 54 percent to 0.2 percent
[Policy] OpenAI published how it detects and disrupts adversarial distillation, where an actor trains a second model on a frontier model's outputs to inherit capability without its safeguards. Its layered defences reduced reasoning-extraction success against GPT-6.2 Astra from 54.0 percent to 0.2 percent. Chain-of-thought traces are the most valuable leak and the hardest to defend. Source: OpenAI
Thinking Machines releases Inkling 2 Mini, a looping open-weights model for laptops
[Models] The 26.4B parameter mixture-of-experts model activates 3.4B parameters and runs its middle layers three times per token, leading open-weights benchmarks below 100B total parameters. At about 20GB it fits consumer hardware, and the team measured no rise in agentic sabotage as loops increased. Capable local models change the data residency conversation for universities. Source: Thinking Machines Lab
Published journal papers contain leftover prompts telling reviewers to be positive
[Education] Inside Higher Ed reports blatant LLM instructions surviving into published papers, the starkest sign yet that authors are trying to game peer review. The failure sits with editors and reviewers as much as with authors. Assessment design that assumes unaided authorship keeps failing in the same way. Source: Inside Higher Ed
Google rebuilt Duplex to call businesses on your behalf inside Gemini
[Tools] Google Research describes how the original booking assistant became a general task agent that dials businesses to gather information and finish errands. Voice agents negotiating in real time cross a usability line that text agents do not. Institutions running phone based services should assume automated callers on the other end. Source: Google Research
A 16-year-old's AI missile startup sells cost per kill to allied militaries
[Geopolitics] Wired profiles Henrijs Dandzbergs, who left high school for Y Combinator and now builds AI-guided munitions through Polemos. The pitch reframes weapons procurement around unit economics rather than capability tiers. Cheap autonomous munitions compress the distance between a software decision and a lethal effect. Source: Wired
Daniel Ek backs Europa Fusion, a fusion startup founded by an AI researcher
[Funding] The Spotify founder's investment moves AI-adjacent capital into energy hardware, the constraint now shaping data center siting. Fusion remains a decade-scale bet, so the signal is about who funds power rather than near-term capacity. Compute and generation are converging into one investment thesis. Source: TechCrunch
University of Chicago hackathon teaches AI skills to IT staff who never wrote code
[Education] The two-day event targeted IT team members with no software development experience, deliberately inverting the usual developer-first training model. Operations staff who can build small tools are the ones who remove manual work. It is the cheapest form of AI capacity building available to a university. Source: EdTech Magazine
Trump threatens a large tax on AI data centers built outside the United States
[Policy] The president raised the idea publicly and said he had not run it past his staff. It lands as hyperscalers weigh overseas capacity against permitting and power constraints at home. A tax on foreign build-out would reprice the geography of AI infrastructure. Source: The Verge
Meta's Muse agent mixed up employees' messages, surfacing balances and orders
[Industry] Employees told TechCrunch that Muse surfaced childcare requests they never made, career advice they never asked for and bank balances belonging to other people. Cross-conversation bleed inside an assistant is a data protection incident, not a bug report. It is the clearest case yet for per-tenant isolation in campus agents. Source: TechCrunch
Google built a mirrored internet to train Chrome agents without live risk
[Tools] Wired reports Google gave its Chrome team a replica internet as a sandbox for browsing agents, so failures do not reach real sites or accounts. Agent evaluation is becoming a question of environment fidelity rather than benchmark scores. Any institution deploying agents needs the same rehearsal space before it needs the incident report. Source: Wired
Language models can rewrite their own weights, researchers report
[Research] The paper describes a model that edits its own parameters after training, instead of relying on a new run, a longer prompt or inference compute that disappears with the task. If it holds, adaptation stops being a separate engineering project. Weight-level self-modification also makes provenance and rollback considerably harder. Source: arXiv
OpenAI releases MedChemBench to test models on early drug discovery reasoning
[Research] The benchmark scores how well models read molecular structures, reason from experimental evidence and pick compounds for early discovery. Domain benchmarks are how a lab decides whether a model is useful rather than merely impressive. Chemistry is a strong test case because wrong answers are expensive and slow to detect. Source: OpenAI
India's Humyn Labs raises 47 million dollars to supply robot training data
[Funding] The company, founded in February, gathers the physical-world data US robotics developers need and draws about 90 percent of revenue from the United States. Data collection is becoming the bottleneck that capital follows. Robotics programs in education will buy from these suppliers rather than build datasets. Source: TechCrunch
The world of AI is evolving at full speed.
Become a Fellow at university-365.com
Become Superhuman... In a world of AI... Prompt Smart, Prompt UP!




























Comments