Google DeepMind’s Agents Find Two Weaknesses in Legacy RSA as Claude Learns to Dream and OpenAI Benchmarks Wet-Lab Science
The late-week AI news was less about new flagship models and more about what agents can now do to the assumptions around them. Google DeepMind said human-guided AI agents found two weaknesses in legacy 1024-bit RSA, while stressing that 2048-bit keys are unaffected. Anthropic published a long engineering essay on agent memory and the “dreaming” passes that let Claude correct and reorganize its notes. OpenAI and Ligo released BioBasicsBench, 100 routine wet-lab tasks on which GPT-6 Astra earned 66.7% of available credit. The Frontier Model Forum published a framework for rating cyber jailbreaks, Apiiro found more than a thousand Claude Code settings files leaking secrets on public GitHub, Visa reported that AI shopping agents failed 42.8% of its benchmark runs, and Meta brought Muse Wonder-trained models to Muse and Meta AI.
Google DeepMind’s AI Agents Find Two Weaknesses in Legacy 1024-Bit RSA
Google DeepMind on Friday published two working papers showing how human-guided AI agents can speed up cryptanalysis, the work of stress-testing the mathematical “hardness” that encryption depends on. In Pushmeet Kohli’s X article “Advancing the frontier of cryptanalysis”, DeepMind’s VP of Science said the team’s agents uncovered two major weaknesses in legacy 1024-bit RSA systems. The research was led by Dj Dvijotham and Abhradeep Thakurta, with cryptographers Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria).
The first result concerns signature forgery when an attacker can feed chosen inputs to a signing device. According to the DeepMind working paper on faster affine roots and a 1024-bit computation, the PKCS#1 v1.5 format offers no extra difficulty once the caller controls more than half a block of data, or a 64-byte digest at 1024 bits. In a sandboxed rerun of a previously solved 1024-bit benchmark from an earlier academic study, the first root took 391 CPU core-years, 3.5 times less than the original computation, and two cooperating AI agents designed, ran and debugged it in under four days. For a shorter 32-byte digest, DeepMind says a new algorithm improves on the previous best method by roughly 1,600× according to theoretical estimates, not an end-to-end run.
The second result, in the DeepMind “Fan Factories” paper on batch number field sieve factoring, is theoretical. With a precomputation costing about one factorization, each further 1024-bit integer could be factored at about 1/58 of the normal cost, 30 times less than Coppersmith’s 1993 multiple number field sieve, at leading order. The paper says a two-agent harness produced the proofs, one model generating and one verifying, with the algebra machine-checked in Lean 4. Kohli was explicit about scope:
“In practice, these results only affect obsolete 1024-bit RSA systems, and modern encryption protocols, which almost universally use 2048-bit or larger keys, remain secure.”
Nobody needs to panic about RSA-2048, but every security team should now assume that hardness margins can move in months rather than decades. Inventory any remaining 1024-bit keys in certificates, SSH and hardware signing services, block untrusted callers from raw signing interfaces, and pull post-quantum migration planning forward instead of waiting for a cleaner deadline.
Anthropic Explains How “Dreaming” Helps Claude Organize Its Memory
Anthropic’s engineering team published Anthropic’s engineering essay on the components of agent memory, describing how Claude stores knowledge between sessions and how offline “dreaming” passes over logs can correct mistakes, write notes, build wikis and decide what gets loaded into future sessions. The idea borrows from sleep-time compute, proposed by researchers at Letta and UC Berkeley. On the Continual Learning Bench database task, Anthropic says Claude Opus 5 was the first of its tested models to close the learning loop, writing rules along with how it knew they were true. In a Pokémon Red comparison about 10,000 steps into the game, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files in two folders plus a top-level global.md and earned four badges.
The dreaming results are concrete. Using about 350,000 internal files, roughly 420 million tokens, from Stainless, the company Anthropic acquired in May, Opus 5 agents built a wiki over 128 sessions; scores on 25 questions rose from 13% to 48% of graded criteria using the wiki alone, and combining wiki and search lifted the score from 54% to 61%, against 53% for search alone, while saving about 20,000 tokens per question. On merger questions that required combining several documents, Claude scored 87% with the wiki versus 73% with search alone. When instructions were rewritten so one agent per contract quoted its terms word for word, the facts that survived in the wiki rose from 41 of 57 to 57 of 57. Anthropic’s steering advice is short:
“We’ve found that it helps to tell Claude what memory is for, but not how to use it.”
This is the most practical memory guidance a frontier lab has published, and it cuts against a lot of current agent design. Stop writing rigid rules about what agents must record, keep the always-loaded memory layer small, and budget scheduled compute to check notes against transcripts, because a confidently wrong memory is worse than none at all.
OpenAI and Ligo Launch BioBasicsBench for Routine Wet-Lab Analysis
OpenAI introduced BioBasicsBench, built with enzyme-design company Ligo, to test whether agents can read the messy evidence of everyday lab work rather than polished figures from papers. OpenAI’s BioBasicsBench announcement on wet-lab agents describes 100 tasks across protein methods, imaging and cytometry, quantification and assays, nucleic acid methods and molecular cloning, built from scientists’ own gel images, colony plates, microscopy stacks, sequencing reads and bench videos. Each task went through three reviews, including one by an external PhD-level wet-lab scientist.
At the xhigh reasoning setting, GPT-6 Astra earned 66.7% of available credit, ahead of GPT-6.1 Sol at 62.0% and GPT-5.6 Sol at 43.8%, while using 4,704 output tokens per task against 9,099 for GPT-5.6 Sol. The evaluation covered seven models and 37 model and effort settings, for 18,500 results. OpenAI highlights cases where Astra spotted a control plate labeled with the wrong antibiotic and recognized a microscope z-stack, but it also documents failures: picking the wrong colonies from a PCR gel, discounting contradictory vacuole images and overcounting a plate by about 12%. The company’s own caveat:
“Higher benchmark scores do not yet mean models can make reliable judgments in the lab.”
BioBasicsBench is useful precisely because it measures boring work, where most lab hours actually go. Biotech teams piloting agents should copy the design: build internal test sets from their own raw data, grade partial credit against expert rubrics, and keep a human checking controls, because a two-thirds score still means roughly one answer in three needs correction.
The Frontier Model Forum Releases a Rating Framework for Cyber Jailbreaks
The Frontier Model Forum released version 1.0 of a framework for rating how dangerous a successful cyber jailbreak is, and OpenAI promoted it on Friday in OpenAI’s X post on the Frontier Model Forum partnership. According to the Frontier Model Forum’s announcement of its jailbreak rating framework, dated November 5, the method scores jailbreaks across four categories: capability gain, capability scope, ease of deployment and ease of discovery. The goal is to stop organizations from rating the same bypass differently and spending scarce effort on low-risk patches before high-risk ones. Amazon Web Services, Anthropic, Google DeepMind, Microsoft and OpenAI executives all gave supporting statements. A public comment period runs through December 21, 2026. OpenAI’s Sasha Baker, Head of National Security Policy, said:
“As AI models become more capable, we need a consistent way to assess jailbreaks that could enable serious cyber harm.”
A shared severity scale is overdue, and it gives enterprise red teams a vocabulary vendors already accept. Map your internal AI security findings onto the four FMF categories now, ask model providers how they triage reports under it, and submit comments before December 21 if your sector’s threat model looks different from the labs’ assumptions.
Apiiro Finds Claude Code “Don’t Ask Again” Files Leaking Secrets on GitHub
Security company Apiiro reported that Claude Code’s “Yes, and don’t ask again” option can save an approved shell command word for word, secrets included, in .claude/settings.local.json. Apiiro’s X post announcing its Claude Code secrets research says it found 2,539 distinct secrets across 1,994 public GitHub repositories. The Hacker News report on Claude Code’s settings file exposure breaks that down: of 104,185 copies of the local settings file in 103,617 repositories, 1,122 held at least one secret, totaling 1,379 different secrets, including 379 database passwords, 43 SSH logins and 23 AWS keys. The larger total adds older file versions, committed shared settings files and saved agent chat logs. Apiiro did not test whether the secrets still worked, though one notified company confirmed its key had been live.
Claude Code adds the file to a user’s global git excludes, but that protection lives on one machine and never reaches the project’s own .gitignore. Common scanners missed many cases: gitleaks flagged 564 of the 1,122 files and TruffleHog 643, mostly because passwords typed into commands do not match known key formats. The researchers recommend rotating every exposed credential first, then running git rm --cached and adding the file to the repository’s ignore list.
Agent folders are the new .env files, and most teams are not scanning them. Add .claude/, .cursor/ and similar directories to pre-commit and CI secret scanning including history, put the local settings file in organization gitignore templates, and train developers to reference secrets as variables so approved commands never capture the value itself.
Visa’s Agentic Commerce Benchmark Finds AI Shoppers Fail 42.8% of Runs
Visa launched the Visa Agentic Commerce Benchmark, which requires an AI model to uncover a shopper’s unstated needs through conversation and then make a correct purchase or no-purchase decision. Per Visa’s press release on its agentic shopping benchmark, the first round tested 22 models from 12 labs on 15 scenarios such as buying a TV, laptop or personal care appliance. The overall pass rate was 57.2%, with the best models at 82.2% and the worst at 28.9%. All three attempts at a scenario passed just 34.2% of the time, and 54% of failed runs still ended with the AI making a purchase. Models that uncovered the most hidden buying factors averaged 70.2%, and the five priciest models per run passed 67.1% versus 50.2% for the cheapest five. Visa plans to extend testing to checkout, payment and order placement. Rubail Birwadker, Visa’s Global Head of Growth, said:
“AI agents have the potential to power the next era of commerce, but trust is key to making it work”
The worrying number is not the pass rate but that most failures still bought something. Retailers and payment teams building agent checkout should require explicit confirmation of discovered requirements before any purchase, log the reasoning trail for disputes, and test consistency across repeated runs, since a single passing demo says little about reliability.
Meta Brings Muse Wonder-Trained Models to Muse and Meta AI
Meta said it is rolling out specialized models trained using its Muse Wonder series to people using Muse and Meta AI. Meta AI Research’s post on specialized models for Muse and Meta AI says the models carry two Muse Wonder capabilities: Active Memory, which tries to tell temporary from permanent information and stale from still-relevant context, and Active Focus, which helps the assistant recover or find alternatives when tasks hit obstacles. Meta reports a 2.4× reduction in forgetting critical information in Muse, a 67% improvement in completing everyday tasks where unexpected issues arise, an Active Memory eval score of 90% versus 76% for the previous model, and a Personal Superintelligence Index score of 47% versus 38%. These are Meta’s own internal evaluations.
Meta and Anthropic shipped memory stories on the same day, which tells you where assistant competition is heading: remembering the right things, not more things. Treat vendor memory metrics as self-reported until independent tests exist, and ask any assistant vendor how users can inspect, correct and delete what the system has decided to keep.
What We Left Out
Anthropic also published a guest science post on AI helping researchers write and revise scientific models as code; we skipped a full section to keep today’s focus on agents and security. NVIDIA promoted Lyra 2.0, its image-to-explorable-3D-world research, but the project page shows the paper and weights first appeared in April, so it is not new. Reports of an unreleased ChatGPT business analytics hub, claims about GPT-6.2 Astra or Fable 5.5 release timing, and data-center financing chatter remain unconfirmed by primary sources and are not reported as fact here.
The through-line this week is that agents are now testing the foundations other systems rely on, from encryption hardness and lab controls to shopping decisions and credential hygiene. Leaders should fund evaluation and guardrail work at the same pace as agent deployment, because the evidence keeps showing capability arriving faster than reliability.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →