AI News Brief: Afternoon roundup, 2026-11-22
Summary
This period was dominated by Anthropic, whose Claude-assisted Riemann hypothesis proof was certified by the Lean and Mathlib communities, and whose IPO process drew attention for its tight controls. Several stories raised the question of whether AI systems can be trusted and measured, from security flaws in cheap inference providers to METR's warning that evaluations are struggling to keep pace. Others showed AI's reach into education, politics, and everyday tools, including a South Korean minister's resignation over exam cheating. Taken together, the stories suggest that AI capability is advancing faster than the testing, security, and governance meant to contain it.
- Polly Hudson describes an experiment in outsourcing conversation to AI, replacing talks with her husband, and found he did not notice. The piece also covers an unwanted AI-related software update on her phone that pushed her to dig into settings.
- A new real-world test places AI coding agents in a simulated developer job with colleagues, project boards, and incomplete handoffs. Even Opus 5.5 and GPT-6 Astra completed fewer than one in ten tasks fully, despite handling clear standalone assignments well.
- Perplexity has launched a Monitor API that runs web searches on a schedule and alerts agents by webhook only when relevant changes occur, using its own model to filter out unimportant results. The launch is part of a push into developer tools as its core chatbot business becomes less distinctive.
- Australian parliament's final four sitting days of 2026 begin Monday, and a joint select committee report is expected to show where Labor stands on AI issues such as datacentres and copyright. The article questions whether the fragile bipartisan consensus on AI will hold through the end-of-year schedule.
- METR reports that leading models including Opus 5.5 and GPT-6 Astra are improving quickly at research tasks but becoming harder to measure. It warns that current data cannot rule out dangerous acceleration, though common tests may also overestimate what the models can actually do.
- South Korea's education minister, Choi Kyo-jin, resigned after an AI-assisted cheating scandal surrounding national university entrance exams. He said he had failed to prepare adequately, following a last-minute scramble to tighten measures.
- This Verge Stepback newsletter examines how vibe-coding, the practice of building software by describing it to AI, is spreading into consumer gadgets. It traces how the trend started and where it may lead.
- Garrison Lovely argues that frontier models can exploit software flaws faster than the world can fix them, and that the risks extend well beyond cybersecurity. He points to recent outages at Microsoft, United Airlines, and JPMorgan Chase as signs of fragility.
- Dragos researchers sent two million fake AI requests to probe the AI supply chain and found at least three compromised inference providers, hundreds of stolen AWS keys, and resellers reading customer requests. OpenRouter responded by removing three providers and 125 endpoints.
- Reuters reports that Anthropic limited its IPO deal teams to four senior bankers at some of its 22 underwriters. It also asked some bank officers to accept six-month restrictions on working with AI rivals, within a compressed and tightly controlled listing timeline.
- Dave Schilling uses the release of Grand Theft Auto VI, and the week's string of tech outages, to explore why video games offer an appealing sense of control. The piece frames gaming as escapism from a world that feels beyond anyone's grasp.
- The Lean and Mathlib communities have officially confirmed Levent Alpöge's proof of the Riemann hypothesis, after dozens of independent verifiers checked roughly 17 million lines of code. The result arrived far sooner than Terence Tao had predicted, and AI did not solve it alone.
- Ken Ono explains why they will not leave mathematics despite AI's progress, arguing that human judgment remains essential. Ono notes that Alpöge announced the Riemann hypothesis proof, made with Claude, on 20 November and that the formal proof has since been certified.
- MarkTechPost compares personal AI agents, including OpenClaw, Hermes Agent, Gemini Spark, OpenAI dots, Meta Muse, and Perplexity Personal Computer, on memory, security defaults, and real cost. The guide is aimed at readers choosing among agents for personal use.
Summaries written with AI from the linked source articles.
Sources in this roundup
| AI (artificial intelligence) | The Guardian | 6 article(s) | |
| The Decoder | 5 article(s) | |
| AI | The Verge | 1 article(s) | |
| MarkTechPost | 1 article(s) | |
| Unite.AI | 1 article(s) |
Most-mentioned keywords
| agents | 2 mention(s) | |
| coding | 2 mention(s) | |
| security | 2 mention(s) | |
| ace | 1 mention(s) | |
| agent | 1 mention(s) | |
| all | 1 mention(s) | |
| alpöge | 1 mention(s) | |
| anthropic | 1 mention(s) |
Sources
- I stopped talking to my husband and let AI do it instead. He didn’t even notice | Polly Hudson
- AI coding agents ace isolated tests but complete fewer than one in ten tasks in a simulated workplace
- Perplexity's new Monitor API turns web search from a one-off query into a nonstop watchdog
- Will fragile bipartisanship on AI survive the silly season, or is parliament heading for a holiday meltdown? | Josh Butler
- AI is getting better faster than experts can reliably test whether new systems are safe, METR warns
- South Korean minister quits after AI cheating scandal engulfs university exams
- The vibe-coding revolution is coming to all the gadgets
- Will AI break the internet? | Garrison Lovely
- Security researchers show how cheap AI inference providers steal keys and snoop on users
- Anthropic Sought Limits on IPO Deal Teams and Work with Rivals
- As the real world falls apart, Grand Theft Auto looks more and more appealing | Dave Schilling
- Lean and Mathlib community officially confirms Levent Alpöge's proof of the Riemann hypothesis is correct
- Why I won’t leave mathematics, despite AI solving centuries-old problems | Ken Ono
- Best Personal AI Agents in 2026: OpenClaw, Hermes Agent, Gemini Spark and More, Compared on Memory, Security and Cost
