DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA, and 2048-bit keys are not affected

😁 Hello, super humans! Last week an outside group measured AI speeding up AI research; this week Google DeepMind points the same trick at cryptography and says its agents found two weaknesses in legacy 1024-bit RSA. We separate what is confirmed from what is still single-source, explain why your 2048-bit keys can relax, and give you a tiny script that shows why key size is the whole game. Also today: a humanoid IPO, a hidden microcontroller inside a LoRa chip, and an induction soldering iron you can build.

📰 Quick Signals

• 🧠 AI: OpenAI reached a reported $1.25 trillion valuation after an $11B round, up from $852B at its $122B round in March.
• 🤖 Robotics: Agility Robotics, the first US-based legged humanoid company to list, closed its first Nasdaq day about 2% above its starting price at $11.49.
• 💻 Programming: Google is giving away gVisor, the code sandbox used by Anthropic and OpenAI, to the open-source community.
• ⚡ Electronics: Semtech’s LR2021 LoRa chip contains an updatable on-board MCU, and Alex Beale reverse-engineered it to run custom firmware.
• 📡 Telecom: RCR Wireless argues NTN-native 6G will feel like one seamless satellite-and-terrestrial service by the early 2030s, though radio standardization limits how identical the two can be.

🔍 The Big Story: DeepMind says AI agents found two weaknesses in legacy 1024-bit RSA, and 2048-bit keys are not affected

If you still have 1024-bit RSA anywhere (old appliances, legacy signing, ancient internal CAs), this is your reminder to find it. If you do not, it is still a clean example of AI agents doing real, checkable mathematics.

What happened: Google DeepMind VP Pushmeet Kohli announced on X that the DeepMind Science team is sharing two cryptographic advances uncovering weaknesses in legacy 1024-bit RSA schemes. Working with leading cryptanalysis experts, the team says AI agents made both advances in just weeks. DeepMind states that keys of 2048 bits and larger are not affected.

The details: The numbers below come from a secondary write-up on DEV Community and have not been checked against the papers, so treat them as claims. It describes two working papers, not yet peer reviewed, titled “Modular Roots from a Restricted Oracle” and “Fan Factories”. The first targets digital-signature forgery: when an attacker controls part of the signed data, a theoretical estimate puts it at about 1,600 times better than the previous best method. The second is batch factorization: by giving each integer its own family of number fields, the next 1024-bit number becomes roughly 30 times cheaper to factor than with the multiple number field sieve under a fixed budget. As a sanity check, the agents reportedly reproduced an earlier 1024-bit computation in 3.2 days on up to 300,000 CPU cores and 16 GPUs, against about five months originally, with proofs written by one agent and verified in Lean 4 by another. Everything was done in an isolated environment and nothing in the reports claims a working attack on a deployed system.

flowchart LR
    A[Cryptanalysis experts set the problem] --> B[AI agents propose attacks and proofs]
    B --> C[Second agent verifies proofs in Lean 4]
    C --> D{Two working papers}
    D --> E[Signature forgery: attacker controls part of signed data]
    D --> F[Batch factorization: cheaper next 1024-bit number]
    E --> G[Legacy 1024-bit RSA weakened]
    F --> G
    G --> H[2048-bit and larger: reported not affected]

🗞️ More News

🧠 AI

• Amazon SageMaker real-time inference now supports NVIDIA multi-instance GPU partitioning, so several models can share one accelerator with hardware isolation, at the cost of about 17 seconds extra provisioning time.
• AMD’s ZenDNN 6.0 and zentorch 2.13 report up to 30% peak and about 10% average FP32 gains over ONNX Runtime 1.17 for BERT-Base on EPYC CPUs.
• WeLM-Image is a unified image-generation model; its Turbo variant uses MeanFlow distillation to render 1024×1024 in 4.98 seconds on one H20 with 8 guidance-free steps.
• OpenAI published BioBasicsBench, a 100-task benchmark for how AI agents handle everyday wet-lab science, from raw instrument data to spotting anomalies.
• OpenAI partnered with Hacktron, a maker of pentest and code-review agents, giving it expanded model access through Daybreak Red.
• Meta’s Muse “Spaces” lets users generate app-like modules on demand, such as shopping hubs, school dashboards and music trackers.
• MIT’s CheMeleon, an open molecular property model built on ChemProp, appeared in 9 of the top 20 entries in the OpenADMET contest.

🤖 Robotics

• Point One’s CEO Aaron Nathan explains on The Robot Report podcast how precise GPS-grade positioning became a platform for robot navigation.
• Pudu Robotics unveiled its “One Brain, Multiple Embodiments” strategy at its European Partner Summit.
• Wing announced a $125M investment, spread over two years, to build its next-generation aircraft in the US.
• Ambi Robotics is installing eight new systems at customer sites ahead of peak season, according to its COO.

💻 Programming

• Cognition turned an open model into a strong coding model but will not release the weights.
• AWS argues developers should stop babysitting their AI agents, pointing to Kiro bots and agent memory.
• Meta’s best AI model now trains the one most of its users get, a look at its specialized Muse models.

⚡ Electronics

• Hackaday spotlights the ESP32-S31, a dual-core RISC-V chip with WiFi 6 and an MMU, running proper Linux, with a 6.x kernel port in progress.
• A receive-only active antenna called Mobile-Whip runs from a single Li-ion cell or USB and integrates a charger, though it is still early-stage.
• Astrolab plans a lunar power and communications network for 2028.
• eeNews Europe interviews Bosch Sensortec CEO Stefan Finkbeiner on sensors, AI and innovation.

📡 Telecom

• Spectrum is overhauling its operating model in a bid to reignite growth.
• Vistance’s CEO is stepping down in the wake of the CommScope dismantling.
• Globe Telcom is expanding its Amdocs deal as it steps up its AI strategy for core infrastructure.

👨‍💻 Code Corner

RSA is only as strong as the cost of factoring its modulus. This script factors a deliberately tiny 65-bit modulus with Pollard’s rho in a blink, then rebuilds the private key to prove it.

import math, random

def pollard_rho(n):
    """Return a non-trivial factor of composite n."""
    if n % 2 == 0:
        return 2
    while True:
        x = y = random.randrange(2, n)
        c = random.randrange(1, n)
        d = 1
        while d == 1:
            x = (x * x + c) % n
            y = (y * y + c) % n
            y = (y * y + c) % n
            d = math.gcd(abs(x - y), n)
        if d != n:
            return d

p, q, e = 4294967311, 4294967357, 65537   # two ~32-bit primes -> 65-bit modulus
n = p * q
f = pollard_rho(n)
g = n // f
d = pow(e, -1, (f - 1) * (g - 1))          # recovered private exponent
msg = 123456789
assert pow(pow(msg, e, n), d, n) == msg
print(f"n has {n.bit_length()} bits, factored as {min(f, g)} x {max(f, g)}")
print("private key recovered, round trip OK")

🧰 Toolbox

• Worktrunk: a Git worktree CLI built for running parallel AI agents without branch collisions.
• agent-desktop: a Vercel Labs CLI for OS-level control through native accessibility trees instead of screenshots, on macOS, Windows and Linux.
• NVIDIA Auto Ontology: a Python framework for graph-based semantic reasoning over enterprise data for LLMs and agents.
• Matt Pocock’s skills: an agent-skills library lifted from his .agents directory.
• RadioThermal soldering iron: an open-source 470 kHz induction soldering station design, with files you can build from.
• Espressif Linux: Espressif’s repository for running Linux on the ESP32-S31.

🛠️ Build of the Week (rotating)

Homemade induction soldering iron: Frank Zhao built two induction irons, one at 13.56 MHz and one at 470 kHz, that heat Curie-point cartridge tips with no temperature sensor at all.

• Difficulty: Advanced
• Parts: an RF power stage and coil, an F-type connector for the 13.56 MHz version, XT30 or USB-PD power input, and a custom handpiece around a commercial tip; the article gives no wattage or part numbers, so follow the linked open designs.
• Why we like it: the cartridge’s Curie point limits its own temperature, so the heater is simple and the heat-up is fast.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: the next step in the video analytics series, after getting clean pixels and moving them over the network, is preparing them for AI.
• Building Your First Neuron From Scratch: follows Part 1’s idea of a network as learnable transformations trained by nudging weights to lower a loss.
• The Network Behind the Cameras: the networking plumbing that moves video once the camera has produced clean pixels.

😀 The Bot Says…

DeepMind’s agents needed weeks to weaken 1024-bit RSA. My own key-rotation policy has been “later” for about the same length of time.

---

That’s all for today! Do you still have a 1024-bit key hiding somewhere? Reply and tell us.