Daily Summary of AI · Live edition, updated continuouslySunday 22 November 2026 · UTC

Latest AI News Today: 22 November 2026

METR warns of AI acceleration as providers leak keys and agents fall short

METR warned that AI research gains may outpace safety testing, with the length of tasks models can handle doubling every 3.6 months and public data unable to rule out drastic acceleration. Security researchers found at least three AI inference providers compromised, with attackers harvesting AWS keys and snooping on user data. A simulated workplace test found that even Opus 5.5 and Astra completed fewer than one in ten tasks completely.

Illustration of the day's news: METR warns of AI acceleration as providers leak keys and agents fall shortIllustration
DS
Daily Summary Of Editors · How we summarise
10 stories10 sources
Image from Illustration: AI coding agents ace isolated tests but rarely finish workplace tasksIllustration
01First reported

AI coding agents ace isolated tests but rarely finish workplace tasks

A new simulated workplace test puts AI coding agents in a developer job with colleagues, project boards and chat histories. Even Opus 5.5 and Astra fully complete fewer than one in ten tasks, but the models are nearing top scores on standard benchmarks.

Image from Illustration: Perplexity adds Monitor API to wake agents on relevant web changesIllustration
02First reported

Perplexity adds Monitor API to wake agents on relevant web changes

Perplexity is expanding its developer business with a Monitor API that watches scheduled web searches and up to 10 pages, sending signed webhooks only when a decision model meets thresholds set by developers. In early internal tests over 60 topics, Perplexity says it delivered 63 fresh relevant pages in eight hours, versus 20 for Firecrawl, 15 for Exa and 10 each for Parallel Core and Lite.

Image from Illustration: METR warns AI research gains may outpace safety testingIllustration
03First reported

METR warns AI research gains may outpace safety testing

METR says AI models are improving quickly on coding and computational tasks, with the length of tasks they can handle doubling every 3.6 months. The organization warns current public data cannot rule out a drastic acceleration in AI research, which could leave humans unable to assess catastrophic risks. It also cautions that common tests may overestimate real world performance because models are trained in similar environments.

Image from Illustration: Lean and Mathlib community confirms Alpöge's Riemann proof is correctIllustration
04First reported

Lean and Mathlib community confirms Alpöge's Riemann proof is correct

The Lean and Mathlib community confirmed that Stanford mathematician Levent Alpöge's proof of the Riemann hypothesis is correct. Alpöge published the proof on Nov. 20, 2026, as 17 million lines of Lean code, and dozens of people independently ran the checks and got the same result. The verification programs are designed so they provably cannot accept a flawed proof. The hypothesis had been one of the most famous unsolved problems in math.

Developing storyAnthropic mathematician claims Claude helped prove Riemann hypothesisFull timeline →
Image from Illustration: Wells Fargo warns customers that use of AI agents is at their own riskIllustration
05First reported

Wells Fargo warns customers that use of AI agents is at their own risk

Wells Fargo will start Tuesday with terms telling customers that use of any AI agent is at their own risk. It joins JPMorgan Chase, Bank of America, Citibank and PNC in new terms or warnings as agents automate logins, purchases and other online tasks. Amazon blocked Meta's Muse from buying on its site, and United says personal bots cannot buy flights. Pushback matters because automation is agents' main appeal, but it can create security risks.

Image from Illustration: Google previews Tool Retrieval in Gemini API for large tool librariesIllustration
06First reported

Google previews Tool Retrieval in Gemini API for large tool libraries

Google has previewed Tool Retrieval in the Gemini API, letting agents keep large tool libraries available without sending every parameter schema in each prompt. The model sees only tool names and short descriptions, then requests full schemas when a task needs a tool, reducing token use from dozens of function declarations. Server-side search is default; client-side mode lets developers run their own lookup, with MCP tools and standard billing.

Image from Illustration: GPT-6.2 Astra and Sol are shorter and less transparent than ClaudeIllustration
07First reported

GPT-6.2 Astra and Sol are shorter and less transparent than Claude

Artificial Analysis found OpenAI's GPT-6.2 Astra and Sol give much shorter, more direct answers than their predecessors, with average lengths falling from 352 to 210 words for Astra and from 337 to 223 for Sol. Their scores for admitting uncertainty dropped sharply, and both trailed Anthropic's Claude Fable 5.2 and Opus 5.5 on a transparency index. The evaluation shows a shift toward brevity and directness while transparency scores lag.

Developing storyOpenAI releases GPT-6.2 Astra and SolFull timeline →
Image from Illustration: Berlin arts complex apologises over military AI eventIllustration
08First reported

Berlin arts complex apologises over military AI event

Kraftwerk Berlin has apologised for a serious error of judgment after transforming its central hall into a war room for the launch of Clarity, an AI targeting system by German arms startup Helsing. The 11 November event featured maps, aerial footage, drones and guards with sniffer dogs, drawing criticism from Berlin's music scene. The row raised questions about the commercialisation of the city's nightlife and its traditional pacifism.

Image from Illustration: South Korea education minister resigns over AI exam cheatingIllustration
09First reported

South Korea education minister resigns over AI exam cheating

South Korea's education minister Choi Kyo-jin resigned Sunday after an AI-assisted cheating scandal before the suneung, the exam that shapes university and career prospects. President Lee Jae Myung accepted his resignation after a top tax accountant exam scorer confessed to using a miniature camera and smartphone for a near-perfect score.

Developing storySouth Korea's AI exam cheating scandalFull timeline →
Image from Illustration: TensorLean verifies GPU kernels and finds 38 correctness bugsIllustration
10First reported

TensorLean verifies GPU kernels and finds 38 correctness bugs

Researchers from Columbia University and CertiK introduced TensorLean, a Lean-based framework that formally verifies GPU kernels at the PTX level. It proves compiled kernels match a tensor-level spec by proving no data races, termination, correct output within tolerance and untouched memory. The team checked 149 kernel instances, including all FlashAttention-2, -3 and -4 kernels plus DeepGEMM, proved 83 correct and found 38 bugs.