Original Geeknewz editorial — synthesizing public reporting from Sep 3–20, 2026. Not a single-outlet rewrite.
On September 19, Anthropic engineer Steve Weis finished a run that sounds like a science-fiction demo and reads like a logistics spreadsheet. With help from Claude, he factored RSA-896—a 270-digit challenge number RSA Security published in the 1990s—by porting the open-source CADO-NFS General Number Field Sieve stack onto GPUs and orchestrating up to 2,048 idle chips for about ten days (~30 GPU-years). OfficeChai’s Sep 20 write-up of Weis’s announcement is careful about the cryptography: no new math, no shortcut algorithm, and no credible threat to today’s 2,048-bit production keys.
That disclaimer is correct. It is also not the Geeknewz story. The story is that coding agents just ate a class of work that used to require a specialist team and a calendar measured in months—and they did it twice in three weeks.
Devin went first; Claude went bigger
Weis’s result landed less than three weeks after Cognition’s Eric Lu published an AI-assisted factorization of RSA-260 (862 bits), completed September 3 with a swarm of Devin agents running a heavily modified, GPU-accelerated CADO-NFS on spare Cognition cluster capacity. Lu’s public notes put the job at roughly 4,900 GPU-days (~13.5 GPU-years) and an estimated ~$400,000 in compute. Before these two campaigns, the public record’s largest comparable RSA challenge factorization was RSA-250 (829 bits) in 2020, which took on the order of 2,700 CPU-years.
Put the two 2026 runs side by side and the pattern is blunt:
- Same algorithmic lineage — GNFS / CADO-NFS, not a magical LLM “break RSA” prompt.
- Same operational trick — accelerate the sieve on GPUs, then treat idle capacity as a fleet.
- Different labs, same month — a coding-agent company and an Anthropic engineer both demonstrated that messy, decades-old scientific codebases can be ported, scheduled, and babysat by agent swarms.
Claude’s quoted credit line, as relayed by OfficeChai, even pointed back at the humans who built the sieve and CADO-NFS. That humility is fine. It does not change the industrial implication: the bottleneck moved from “who can rewrite HPC Fortran” to “who can point agents at unused silicon.”
The same weekend, Nvidia sold “data centers in the house”
While the factoring notes circulated, Tom’s Guide published a Sep 20 interview with Nvidia AI product leader Adel el Hallak arguing that “local AI is here.” El Hallak’s home already runs a DGX Spark ($4,699) overnight; Perplexity’s Portable Computer package sits on the same hardware with a local-first harness; and RTX Spark laptops with GB10 silicon and 128GB unified memory are lined up from Dell, HP, Lenovo, Asus, MSI, Acer, plus a Microsoft Surface Laptop Ultra—expected this fall, running Windows 11 rather than Linux-only appliances.
Geeknewz is not claiming your next gaming laptop will crack RSA-896 over a weekend. We are claiming the mental model just flipped. Cloud chatbots taught consumers to rent intelligence by the token. Nvidia’s pitch—and the Claude/Devin factoring demos—teach engineers that owned or scavenged GPUs plus an agent harness can execute multi-day scientific campaigns that previously needed institutional ops.
That is why el Hallak’s line about IT departments suddenly becoming “cost optimizers” after token bills arrive matters next to Weis’s idle-GPU fleets. The industry is simultaneously:
- selling subscriptions for frontier models in the cloud,
- shipping personal AI appliances that keep files home, and
- proving agents can commandeer spare training silicon for classic cryptography workloads.
What this is—and what it is not
It is not “AI broke the internet’s encryption.” Weis and Lu are explicit that deployed key sizes remain astronomically out of reach; the cost curve still climbs steeply with bit length. Banks and browsers do not need a panic patch because Claude ported CADO-NFS.
It is a capability milestone for agentic software engineering:
- Porting brittle, specialist HPC packages across architectures.
- Orchestrating thousands of devices and scavenged capacity.
- Closing the loop from code change → job schedule → verified mathematical result.
That loop is the same shape as the containment and verification debates Geeknewz covered earlier this week—Gemini’s real-world logins during cyber tests, military hallucination near-misses, Muse-style desktop agents with mail and file keys. When agents can run multi-day compute campaigns, “did the model stay in the sandbox?” stops being a lab curiosity and becomes an ops question for whoever owns the GPUs.
Geeknewz take
September 2026’s RSA challenge headlines will be misread as crypto doom. Read them as a labor-market and infrastructure signal instead. Coding agents compressed a niche HPC workflow into scavenged GPU-years. Nvidia, in the same news cycle, is normalizing personal AI computers as household appliances. Frontier labs keep warning about escape and oversight while their own tools demonstrate how to weaponize idle silicon for legitimate science.
The uncomfortable Geeknewz conclusion: the frontier is no longer only about bigger chat models—it is about who controls the agent that wakes up the GPUs at 2 a.m. Whether those GPUs sit in a Cognition cluster, an Anthropic engineer’s scavenged pool, or a $4,699 box humming under a desk, the cryptography challenge numbers were just the receipt.