Claude cracks NIST curve seeds & Anthropic Public Benefit Fund launch - AI News (Oct 20, 2026)

For decades, cryptographers have wondered whether the numbers behind some of the world's most widely used encryption standards hid a secret backdoor. Now an AI says it has found the real answer, and it is far more ordinary than anyone expected. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 20th, 2026. We've got a busy one: Anthropic's numbers on agents at work, a new open model from DeepSeek, a leaked image model from OpenAI, and an update on that AI-powered hacking spree.

Claude cracks NIST curve seeds

Let's start with that cryptography mystery. The elliptic curves standardized by NIST, the ones with names like P-256, were built from published seed values whose origins were never explained. That gap fueled years of suspicion that someone might have chosen them to weaken the math. Anthropic says Claude traced those seeds back to an old C library random number generator, seeded with Unix timestamps. No hidden hash, no elaborate story. Anthropic stresses this is not a break of the cryptography, and if anything it makes backdoor theories less believable, since the process looks simple and reproducible. A few loose ends remain, including one unexplained seed, but a long-running cloud over these standards just got a lot thinner.

Anthropic Public Benefit Fund launch

Staying with Anthropic, the company released some striking internal data. By August of this year, hours worked by AI agents inside the company outnumbered employee hours six to one. Hiring for junior technical roles has slowed, and the company now looks more for autonomy and judgment than narrow expertise. Productivity is up across engineering, sales, legal and finance, but Anthropic says the real bottleneck has shifted to human attention, since staff spend so much time reviewing what agents produce.

AI agents outnumber Anthropic staff

That bottleneck is exactly what the Claude Code team tackled in a separate post about code review. Rather than having people read every change, they push checks earlier and let Claude verify work, review pull requests, and flag security issues, saving human approval for riskier changes. From June to September, merged pull requests more than doubled while reverts dropped by over 40 percent. A widely shared post on X makes a similar case from another angle: if you want to trust agent-written code, invest in strong architecture, types, and tests, so the correct path is the default, rather than piling on human reviewers who simply don't scale.

Rethinking code review for AI

One more Anthropic item. The company is creating a Public Benefit Fund, initially earmarking 5 percent of its shares, pending stockholder approval, with more added as its valuation climbs. The money could go to nonprofits or directly to groups like workers displaced by AI. It's structured as a company-owned LLC rather than a charity, and it can't be used for lobbying or political donations. Combined with earlier pledges, Anthropic says about a fifth of its shares are now dedicated to the public good.

AI agent hacking campaign update

Now an update on the AI agent hacking story. The Verge reports the campaign hit more than 100 websites, with victims including John Deere, SharkNinja, and Cummins. The attacks were tied to exposed instances of Cairn, an open-source agent orchestration platform, targeting firms in the U.S., Canada, Australia, and Japan. The companies say they're investigating or have contained the issue, and John Deere says it sees no sign sensitive data was accessed.

DeepSeek V4.1 Flash update

On the model front, DeepSeek shipped a minor but notable update called V4.1-Flash-1017. It's a large open-weight model under the MIT license, claiming solid gains in coding, agentic work, and vision, while also finishing tasks faster with shorter outputs. DeepSeek also documented built-in web Search and Fetch tools, letting developers give agents live web access without setting up their own search infrastructure.

OpenAI Aurora image model leak

OpenAI, meanwhile, may have tipped its hand. Three image models codenamed Aurora briefly showed up in its API over the weekend and then in blind arena tests before vanishing. Many suspect this is GPT-Image-3. Early impressions suggest one variant trades quality for speed, another the reverse. No official word yet.

Replit switches default to OpenAI

And alliances keep shifting. Replit's CEO Amjad Masad said at TechCrunch Disrupt that OpenAI is now Replit's default model provider, citing cost, partnership, and performance. Users can still pick Claude, but most stick with defaults, which could mean a real dent for Anthropic, since Replit was one of its biggest customers. People close to the deal dispute that the relationship is over entirely.

Quasi-Riemann proof holds up

In mathematics, a follow-up to OpenAI's proof of the quasi-Riemann hypothesis. Quanta reports mathematicians have now dug through it far enough to believe it's correct. The result narrows where the key zeros of the zeta function can sit, sharpening what we can say about prime numbers. Experts call it a clever combination of older ideas rather than brand-new math, and caution that the full Riemann hypothesis is still a long way off.

Qualcomm on 100B-parameter phones

Over to hardware. Qualcomm CEO Cristiano Amon says some leading AI companies want phones that can run a 100-billion-parameter model on the device itself by 2028, and some are building their own phones, like ByteDance's Doubao handset. His argument: phones hold the personal context and apps an agent needs to act for you. He also warned that rising memory prices are squeezing the market, especially cheaper phones, and teased 6G demos at the 2028 Los Angeles Olympics.

AI tools de-bloat smart TVs

Finally, two stories about people taking control back. Frustrated smart TV owners are using AI coding tools to strip out ads, tracking, and sluggish home screens, sharing projects on GitHub and viral posts about cleaning up their sets with Claude Code. Results vary, and some folks have bricked their TVs, but if this spreads, it could worry the smart TV ad business. And one developer fine-tuned a small Qwen model on a MacBook Air to unlearn AI writing habits, the em dashes and tidy lists of three. The output is a bit flat and still hallucinates at times, but it shows style can be trained in, not just prompted.

That's it for today's Automated Daily, AI News edition. Links to all the stories we covered can be found in the episode notes. I'm TrendTeller. Thanks for listening, and see you tomorrow.

Claude cracks NIST curve seeds & Anthropic Public Benefit Fund launch - AI News (Oct 20, 2026)

Our Sponsors

Today's AI News Topics