- DeepMind's AI agents turned up two weaknesses in legacy 1024-bit RSA, with a claimed 1,600x speedup on signature forgery and roughly 30x cheaper bulk factoring. Modern 2048-bit keys are not affected.
- Gemini Omni 1.2 Flash reached general availability with 15-second video clips, up from 10, and a 30% price cut to about $0.07 per second at 720p.
- An independent assessment puts Anthropic's rate of capability improvement at roughly twice its 2024 to 2025 pace, then says Anthropic redacted most of the evidence behind that number.
- The Justice Department stood up a Super Intelligence Task Force on November 5 to prosecute misuse of advanced AI by foreign states, criminal networks and companies.
- On the open side, NVIDIA published downloadable Cosmos 3 world models and Baidu released a 280-billion-parameter base model with no instruction tuning at all.
- Today's thread is verification, which is also the question underneath the AI automation tool landscape.
Every headline move today came with an asterisk about who checked it. A cryptanalysis result where the big numbers are theoretical. An independent audit of a frontier lab where the lab held the red pen.
The capability is real in every case, and the evidence is partial in every case. Which points at a small, boring habit: the number that decides anything for you is the one you measure.
The Front Page: DeepMind's agents found two cracks in retired encryption
Google DeepMind says AI agents, supervised by humans and working with academic cryptographers, found two previously unknown weaknesses in 1024-bit RSA. RSA is the math behind a lot of older digital signatures and key exchange, and the number is the key length: 1024-bit keys left serious use years ago. Pushmeet Kohli, DeepMind's VP of science, announced the work on November 6, and Interesting Engineering has the clearest writeup. Nadia Heninger of UC San Diego and Emmanuel Thomé of Inria were on the team.
The first weakness concerns padding, the extra structure wrapped around a message before signing to make forgery harder. Once an attacker controls about half the signed block, the team says, padding buys almost nothing, and where the attacker controls only a short fingerprint the agents produced an algorithm DeepMind estimates is around 1,600 times faster than the best published method. The second is batch factoring: when many keys share a bit length, one expensive precomputation can be spread across all of them, cutting the average cost per additional key by roughly 30 times.
Now the asterisks. Both multipliers are theoretical rather than measured end to end. The part that ran was a sandboxed reproduction against a 1024-bit key an academic team had already broken, published on the IACR ePrint archive in September, and there DeepMind reports its agents used 3.5 times less CPU time than factoring the key outright. No production system still relying on 1024-bit RSA has been named, and DeepMind says modern encryption is untouched.
What it means: Encryption did not fall over. What moved is that the slow, intuition-heavy part of cryptanalysis is now something a machine can take a real run at, under supervision, in weeks. That shortens the honest answer to how long any given assumption keeps holding. The job this week is an inventory, not a panic.
Releases & Features
Gemini Omni 1.2 Flash. Google's video model is generally available, and the change that matters is price, not length. Clips run up to 15 seconds instead of 10, and Google says pricing is 30% below the previous version, which third-party reporting puts near $0.07 per second at 720p. Google's announcement. It debuted top of the Arena image-to-video leaderboard at 1502, a crowd preference ranking rather than a measure of whether it renders your product right.
NVIDIA's Cosmos 3 world models. The smaller move and the quietly useful one. NVIDIA put open-weight checkpoints for interactive world simulation on Hugging Face on November 5, for driving and robotics. A world model generates what a camera or lidar would see next given an action, so a robot policy trains against a simulation instead of a parking lot. Model cards here, developer writeup here. The smallest configuration runs on one workstation GPU.
Inkling 2 Small. Thinking Machines Lab released a 594-billion-parameter mixture-of-experts model, meaning only a slice of the network runs on any query, with roughly 17 billion parameters active per request. Weights stay closed during external safety testing, then follow under Apache 2.0. Details, with third-party evaluation.
What it means: Two kinds of cheap arrived today. Google cut the price of a hosted capability by a third, which makes a workflow you shelved six months ago worth another look. NVIDIA gave away weights that run on hardware a small team can buy. The first is faster to adopt. The second nobody can take back.
In the Lab
METR, an independent evaluation nonprofit, published an assessment on November 5 of how much AI is speeding up AI research inside Anthropic, based on on-site work from August 12 to October 29. Its estimate: Anthropic is improving model capabilities on well-specified tasks at about twice the rate it managed from March 2024 to October 2025, a trend of 1.9x with a 1.5x to 2.8x interval. Roughly 80% of that is attributed to AI itself, which METR calls a subjective judgment. The full assessment.
The sharper part is what you cannot read. Anthropic held unilateral redaction rights and cut a majority of the evidence METR found most informative, calling it material that could help rival developers. A planned check on how Anthropic splits compute between capability and safety work was canceled partway through.
What it means: This is the most careful outside look anyone has had at whether AI is compounding on itself inside a frontier lab, and it reaches you with the workings removed. Treat the 2x as a serious estimate, not a measurement. The evaluator with the most access published what it was allowed to publish, which is roughly the current ceiling on independent verification.
The Oversight Desk
The Justice Department launched a Super Intelligence Task Force on November 5, led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss. The announcement frames the work around foreign state actors, transnational criminal organizations, and fraud aimed at taxpayers and critical infrastructure. Weiss said the department wants companies to exercise "vigorous internal controls so that SI technology behaves as intended." Nextgov has the structure here.
What it means: Enforcement, not rulemaking. No new obligation lands on anyone, and nothing here creates a standard to comply with. What it creates is a prosecutorial interest in whether a company that shipped a capable system had controls around it, answered later from your logs by an unsympathetic reader. If you run agents that act in the world, the cheap insurance is an audit trail you would be comfortable handing over.
Nobody is going to verify a vendor's numbers for you. The test that counts runs on your own data, on a schedule, and tells you when something slips. Describe what you rely on and get the spec back.
On the Radar
Four smaller moves, sources attached.
- Baidu shipped a base model on purpose. ERNIE 5.5 Flash Base is a 280-billion-parameter model with 14 billion active per input, trained on 33.5 trillion tokens and published with no instruction tuning so teams post-train it themselves. Source.
- The first humanoid robot stock. Agility Robotics began trading on Nasdaq as AGLT on November 5 at roughly a $2.5 billion pre-money valuation, after a SPAC merger that brought in about $465 million. Reported first-half 2026 revenue was $174,000, mostly from related parties. Source.
- Perplexity put agents on a clock. Teammates are always-on work agents that keep running and monitoring rather than waiting in a chat window, for Max subscribers on the web. Source.
- A decade of Earth, searchable in plain English. Planet released a free app for querying roughly ten years of satellite observations in natural language, a niche tool that quietly takes a specialist out of the loop. Source.
The Bottom Line
The gap between a capable claim and a checkable one is widening. DeepMind's agents did work that used to need a career of intuition, and the proof arrives as an estimate. METR got closer to a frontier lab's internals than anyone and published around the redactions. Watch whether the cryptanalysis result turns up as a paper with reproducible numbers, the difference between a milestone and a press cycle. Meanwhile the leverage sits where it always does, in picking one job you repeat weekly and getting something running on it you can inspect yourself.
Frequently Asked Questions
-
No. The findings apply to 1024-bit RSA, a key size retired from serious use years ago, and DeepMind says modern 2048-bit keys are unaffected. The practical takeaway is an inventory job rather than an emergency: find out whether anything you run still signs or exchanges keys with a 1024-bit certificate, which usually turns up in old internal services and embedded hardware rather than anything public. The broader pattern, machines doing the specialist work that used to gate a field, is the shift described in where generative AI turns into functional AI.
-
Check who ran the test and whether anyone outside has reproduced it. A score a company reports on its own benchmark is a claim, not a result, and a leaderboard rank measures what raters preferred rather than what your job needs. The only number that settles anything is the one you get running the thing on your own data, including examples it has never seen. Building that check into a repeatable step is easier than it sounds, and the workflow directory has patterns you can copy.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
