· 12 min read
Claude's memory dreams; wet-lab agents and RSA-1024 attacks
Published
Anthropic explains Claude's memory dreaming loop, OpenAI introduces BioBasicsBench, and DeepMind agents accelerate attacks on obsolete RSA-1024.
Today in 30 seconds
- Anthropic: Claude’s memory has a “dreaming” loop
- OpenAI + Ligo ship BioBasicsBench for everyday wet-lab agents
- DeepMind: AI agents accelerate cryptanalysis on legacy RSA-1024
- Frontier Model Forum Jailbreak Rating Framework v1.0 for cyber threats
- Meta brings Muse Wonder–trained specialized models (Active Memory / Focus) to Muse + Meta AI
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Pour the tea: overnight Anthropic explained how Claude “dreams” to fix and reorganize agent memory between sessions, OpenAI put wet-lab agents on a 100-task bench of messy gel plates, and DeepMind’s human-guided agents found faster attacks on legacy RSA-1024. xAI’s API text watermark also flips on by default today.
Top news
Anthropic: Claude’s memory has a “dreaming” loop
On 6 Nov Anthropic published The components of agent memory: agents should capture salience in-session, then use offline “dreaming” (sleep-time compute over transcripts) to correct mistakes, build wiki-style maps, and promote a tiny injected context layer for the next session.
- Dreaming is framed as hippocampus-style consolidation: Claude replays logs between sessions to rewrite notes, build linked wiki pages, and float high-value rules into preemptively loaded context (Claude Code’s auto-memory header stays small: first 200 lines / 25 KB of
MEMORY.md). - On Continual Learning Bench 1.0, stronger Claudes close a learning loop (record failures → test theories → write verified rules).
- On Pokémon Red, Fable 5 kept 9 files and earned 4 badges vs Sonnet 3.7’s 413 files / 225 folders and one badge — salience judgment beats archival dumps.
- Dreaming corrected a wrong Mt. Moon landmark note against the transcript and changed the run outcome .
- On a ~350k-file / ~420M-token Stainless corpus, scaling dreaming from 2→128 sessions raised graded Q&A from 13% → 48%; wiki + search beat search alone (61% vs 53% aggregate) and saved ~20k tokens/question.
- Goal-steered dreaming on a Harvey LAB contracts task moved graded facts in the wiki from 41/57 → 57/57.
- Managed Agents also expose Dreams as a research-preview API (
dreaming-2026-04-21beta): async job over a memory store + 1–100 sessions that outputs a new store without mutating the input.
Why it matters: Persistent agents are becoming a systems problem (when to write, how to repair, what to inject), not a bigger context window — and Anthropic is documenting the evals that actually move the needle.
- Do this now: Read the engineering post; if you run Managed Agents, try a dream on a messy store with narrow
instructionsbefore trusting injected context.
Sources: Anthropic engineering (6 Nov) · Managed Agents Dreams docs · @AnthropicAI
Get this brief in your inbox every morning — free
OpenAI + Ligo ship BioBasicsBench for everyday wet-lab agents
OpenAI’s 6 Nov post introduces BioBasicsBench — 100 routine wet-lab tasks built with enzyme-design firm Ligo from raw gels, plates, microscopy, FCS, FASTQs and protocol video — scoring whether agents interpret evidence, catch bad controls, and refuse unsupported conclusions.
- Domains: protein methods, imaging/cytometry, quantification/assays, nucleic acids, molecular cloning. Triple review (Ligo author peer + external PhD wet-lab + final internal).
- At xhigh reasoning: GPT-6 Astra 66.7% mean reward (CI 60.3–73.1; ~4.7k out tokens/task) vs GPT-5.6 Sol 43.8% (~9.1k tokens). GPT-6.1 Sol 62.0%. Evaluation: 7 models × effort settings, 5 attempts × 100 tasks (18,500 graded runs).
- Failure modes still common even for Astra: assumption overriding controls on colony PCR, generalizing vacuole contrast from one focus plane, ~12% colony overcount on a plate.
- Framing is dual-use aware: better lab copilots accelerate biology and need biosafety tracking.
Why it matters: Paper-figure benchmarks hide the ambiguity scientists live in; this is the first major lab-agent eval aimed at that mess.
- Do this now: Skim the Astra vs Sol qualitative examples before trusting any “lab agent” demo.
Sources: OpenAI (6 Nov) · @OpenAI
DeepMind: AI agents accelerate cryptanalysis on legacy RSA-1024
Human-guided DeepMind agents uncovered two weaknesses on obsolete 1024-bit RSA proving grounds — a ~1,600× better estimated signature-forgery algorithm in the low-control regime, and a batch-factoring method ~30× cheaper per additional same-size key — while modern 2048-bit+ RSA remains the intended safe zone.
- Shared 6 Nov via @GoogleDeepMind / Pushmeet; technical PDFs: Faster Affine Root… and Fan Factories… on DeepMind media storage. Collaboration includes Nadia Heninger (UCSD/Inria) and Emmanuel Thomé (Inria).
- Signature result: once an attacker controls roughly half or more of a padded block, padding adds no forgery protection beyond the no-padding case.
- Sandboxed replay on a published 1024-bit benchmark key used 3.5× less CPU and <0.1% of the compute to fully break the key.
- Point is methodological: AI shortens the historical years-to-decades cadence of hardness-assumption stress tests — defenders should treat hardness as something to re-probe, not a fixed constant.
- DeepMind stresses live modern protocols with ≥2048-bit keys are not claimed broken by these results.
Why it matters: Same agent stack that finds software vulns is now chewing math assumptions; crypto standards bodies need faster re-evaluation loops.
Sources: @GoogleDeepMind (6 Nov) · Pushmeet article / papers linked in thread · PDF 1 · PDF 2
Frontier Model Forum Jailbreak Rating Framework v1.0 for cyber threats
FMF (5 Nov; amplified by OpenAI 6 Nov) released a preliminary shared method to rate cyber jailbreaks across four axes — capability gain, capability scope, ease of deployment, ease of discovery — into Negligible→Critical-style bands so labs stop treating every bypass as equal.
- Goal: comparable severity so teams patch high-risk unlocks first. Quotes of support from AWS, Anthropic, DeepMind, Microsoft, OpenAI security/policy leads.
- Version 1.0 is explicitly preliminary; public comments open through 21 Dec 2026 (
[email protected]). - Sits next to OpenAI’s ongoing Daybreak / GPT-6 Cyber work (already covered 5 Nov) as process infrastructure rather than a new model drop.
Why it matters: Without a common scale, “we jailbroke the model” is marketing noise; this is the industry trying to grade the harm.
Sources: FMF update (5 Nov) · @OpenAI
Meta brings Muse Wonder–trained specialized models (Active Memory / Focus) to Muse + Meta AI
Meta (6 Nov) is shipping specialized versions of existing Muse/Meta AI models trained with Muse Wonder, pushing Active Memory and Active Focus to general users — not only the frontier Wonder series.
- Personal Superintelligence Index: current Muse 47% vs previous 38%.
- Active Memory eval: tasks correctly using critical context 90% vs 76%; Meta claims 2.4× fewer critical forgets for goal help.
- Active Focus: 67% improvement completing everyday tasks when unexpected issues appear; models better multitask, recover, and accept mid-run steers.
- Pitch: “personal superintelligence” capabilities distilled into the models people already open every day.
Why it matters: Consumer assistants are getting the same memory/focus scaffolding labs hyped for agents — expectation bar for “it forgot again” just moved.
Sources: Meta AI Research (6 Nov) · @AIatMeta
xAI API text watermarking turns on by default today (7 Nov)
Per xAI docs, starting 7 November 2026 text outputs from listed Grok models are invisibly watermarked by default (EU AI Act framing); check x-watermark response header; team admins can opt out per model in Console.
- Models listed: grok-4.7, 4.6, 4.5, 4.3, grok-4.20-reasoning / non-reasoning / multi-agent. Applies to Responses + Chat Completions, including structured outputs and tool-call arguments.
- Watermark shapes token choice (no extra characters); detection is statistical and gated by application. Docs claim no quality/speed hit.
- Opt-out is team-wide per model, audited; some models may be locked on. Changes apply within ~5 minutes.
Why it matters: If you pipe Grok into pipelines that assume unmarked text (training sets, provenance tools, compliance), today is the cutover — verify headers and opt-out policy now.
- Do this now: Log
x-watermarkon a canary call; decide which models stay on for EU-facing traffic.
Sources: xAI watermarking docs
Visa Agentic Commerce Benchmark: shopping agents pass ~57% of runs
Visa’s 5 Nov benchmark forces 22 models from 12 labs through 15 realistic shop scenarios (TV, laptop, etc.): uncover hidden requirements via dialogue, then correctly buy or refuse. Overall pass 57.2%; 42.8% of runs fail — and 54% of failures still end in a purchase.
- Best models 82.2% pass; worst 28.9%. Models that uncovered the most hidden factors averaged 70.2% vs 44.4% for the least curious quintile.
- Consistency: all three attempts on a scenario passed only 34.2% of the time. Higher-cost models did better on average (67.1% vs 50.2%) but didn’t guarantee wins.
- Visa Trust Index (Sep 2026): 72% of US consumers used an AI assistant; only 23% trust AI to handle payment. Interactive materials on Visa Payments Frontier; more scenarios (checkout/payment) planned.
Why it matters: Agentic checkout demos are ahead of reliability; “almost bought the wrong TV” is the failure mode that kills trust.
Sources: Visa press release (5 Nov)
Gemini Omni 1.2 Flash: native 15s clips, multi-turn edits, −30% price
Google DeepMind (5 Nov, GA on API/Cloud) shipped Omni 1.2 Flash with 3–15s native video (was 10s), conversational multi-turn edits, stronger reference adherence, and $0.07/s at 720p (30% below 1.1). Outputs carry SynthID; Provisioned Throughput available.
- Workflow: draft cheap at 360/720p → edit without full regen → upscale winners to 1080p/4K. Better lip-sync, multi-speaker conditioning, continuous camera moves with fewer jump cuts.
- Surfaces: Gemini API / AI Studio, Google Cloud, Flow, Google Vids (15s at no extra cost with account/Workspace), Gemini app rolling to AI Ultra, Google Ads.
- Positioned for ads, try-on, catalog spins, explainers — still weaker on fast sports / complex physics per Google’s own caveats.
Why it matters: Cheap, editable 15s is the unit commerce and social actually buy; this is Google pushing Omni from preview into production unit economics.
Sources: Google blog (5 Nov) · Model card
Anthropic: teaching AI to build scientific models
Anthropic’s 6 Nov research post Teaching AI to build scientific models (Science Blog) describes agents that propose candidate models, fit parameters, and revise equations — and, more importantly, return a distribution of plausible models so scientists see which features the data actually support.
- Framing: modeling help beyond chatbot Q&A — generate alternatives, expose under-determined structure, point to experiments that would discriminate.
- Honest limit stated in the piece: current systems have mostly rediscovered known models; solving genuinely open problems remains limited.
- Same-day @AnthropicAI post pointed readers here alongside the dreaming/memory engineering work — twin “science + agents” narrative during IPO season.
- Pair with BioBasicsBench: OpenAI measuring messy wet-lab vision; Anthropic narrating equation/model search.
Why it matters: Separates “AI wrote a paragraph about biology” from “AI proposes falsifiable model families” — and marks how far the latter still has to go.
Sources: Anthropic research · @AnthropicAI
NVIDIA Build free API access to Moonshot Kimi K3.1 Pro (+ peers)
X News + builder chatter on 6 Nov: NVIDIA’s Build platform is offering no-credit-card API access to Kimi K3.1 Pro (Moonshot, ~2.8–3T-class MoE, 1M context, text+image; weights on HF since ~4 Nov), plus peers such as DeepSeek-V4.1-Flash and GLM 5.3 Flash — useful try-before-you-buy for agentic coding stacks.
- Treat vendor/X-summary numbers as secondary until you hit primary Moonshot/NVIDIA cards for exact active-param and license terms; open weights already on Hugging Face per coverage.
- Cerebras/Akash demos circulated (e.g. fast purchase flows); local self-host still needs multi-TB class hardware for full MoE.
- Fits the week’s pattern: western open MoE (Inkling 2 Small, covered yesterday) plus Chinese open giants getting free inference funnels.
Why it matters: Free frontier-class endpoints collapse the “can I even try it?” tax for indie agent builders this weekend.
Sources: X News story · X News — Kimi K3.1 Pro · build.nvidia.com
GitHub repos
- View repo →
datalab-to/indra
~1M-param OCR aimed at CPU; claims ~15× Tesseract speed with accuracy near much larger models. Clear primary artifact for doc pipelines.
- View repo →
hevnsnt/CVEngine
AI-first external attack-surface assessment with continuous discovery and evidence-based vuln validation (defensive/red-team adjacent).
- View repo →
gh9ko8y/graphmemory
Bitemporal agent memory (decay, TTL, competitive retrieval on Neo4j+) — thematic twin to today’s Anthropic dreaming piece.
- Honorable weekly trending (already large): calesthio/OpenMontage agentic video production system — note high star count; use only if you need the montage stack.
Research papers
- Read paper →
EvoMemState (2611.03743)
— state-centric personal-agent memory (Relational + Reasoning states); community note ~57.6% on MemSyco-Bench (+~12 pts) with a separate GRPO memory-policy model.
- Read paper →
Harness-S (2611.04277)
— train with sibling harnesses that rescue failed RL rollouts, then transfer gains back to the original harness (AppWorld Pass@3 example 64.89 → 72.52 on Qwen3.5-9B per author thread).
- Read paper →
SLA (2611.08373)
— Scaling Spatial-Language-Action via continual pre-training.
- Read paper →
VFlash (2611.08449)
— fast visual planning via flash primitive decoding.
- Read paper →
VisSearch-Bench (2611.04894)
— long-horizon visual search for multimodal agents.
AI workflows
1. A construction company president’s Grok Bot playbook: ~40 hours a week of ops, draft-first · Workflow page
- Replaces: the owner personally chasing email, bidding, invoices, pay apps and house admin, and rolling the same habits out to Ops, Admin and Safety directors by hand.
- After about two months on Grok Bot (since late August), the president estimates ~40 hours a week of equivalent work; directors got bots in the last two weeks.
- A Chief of Staff bot also screens viral “you must do this” posts and calls ~80% engagement bait.
- What to connect: Grok Bot with one named bot per job (email, bidding, invoices, pay apps, house), email and calendar connectors, routines for morning digest, promise keeper, Friday look-ahead and bill nudges .
- Skills for new-job folders, subcontract intake and invoice review.
- Voice skill trained from real Sent mail.
- Drafts only until you press Send; nothing paid, deleted or sent without OK.
- Usage kept down with shorter routines and pausing one-off reminders after they fire.
- Who it’s for: owners of small construction or trade businesses (and similar sole-prop ops) who want bots on real jobs without letting them send or pay alone.
2. One podcast drop becomes 23 posts on 6 platforms in about 3 minutes · Workflow page
- Replaces: half a day of cutting clips, writing an X thread, LinkedIn posts, a newsletter, SEO articles and scheduling them by hand after each long recording.
- A 48-minute episode (9,214 words transcribed, 8 scored moments kept) became 8 shorts, 4 X threads, 4 LinkedIn posts, 4 Instagram quote carousels, 1 newsletter, 2 SEO articles and a week of scheduled slots .
- The author claims ~11 hours saved and a ~3-minute run once the file lands.
- What to connect: Grok Bot watching a
/videosfolder (or equivalent drop trigger), then parallel specialist agents for transcription/moment scoring, clip cutting, thread writing, LinkedIn, quote cards, newsletter, blog and scheduling.- Peak-time slots named (9 AM, lunch, 5 PM, 8 PM).
- Prompts for each agent are in the author’s follow-up; keep a human OK before anything posts if you treat the calendar as drafts.
- Who it’s for: newsletter and multi-platform creators who already record long-form audio or video and want one recording to fill a week.
3. Weekly “Bot Hours Saved” audit: which Grok Bots actually ship work · Workflow page
- Replaces: reading every bot chat to guess what’s useful, or trusting daily status pings that don’t prove shipped work.
- One Friday bot reads seven days of conversations and routine runs, counts only concrete completed work, and ranks the fleet.
- This week’s estimate: ~183 hours of equivalent human effort, including ~133 hours tied to live output (assumptions stated; undocumented work can be missed).
- What to connect: a Grok Bot with read access to the other bots’ chats and routine history, one durable Page for the weekly report, and a Friday routine that posts only headline numbers plus a link in chat.
- It ranks bots, flags idle and overlapping jobs, lists work waiting on approval, and charts the weekly trend; it recommends pauses or merges and does not change other bots itself.
- Author offers a shareable template (reply HOURS on the post).
- Who it’s for: people running several Grok Bots on a fixed weekly plan who need a honest weekly scoreboard of hours saved versus chatter.
Quick hits
-
FMF + OpenAI cyber jailbreak process is live for comments — if you file red-team findings, map them to the four FMF axes before Dec 21 so your severity language matches member labs. FMF
-
@swyx commits to weekly Anthropic IPO coverage with @jaminball until it prices or doesn’t — useful running commentary track beside the S-1. Post
-
Harness engineering keeps eating timelines — X News roundup of Cherny-style prompt/budget/system scaffolds and $500k–$1.2M “harness” hiring; pair with today’s Harness-S paper in HF Papers. Story
-
OpenAI Business Hub analytics spotted — unverified-in-product chatter that verified businesses may soon see ChatGPT mention impressions/clicks (Search Console for answers). Treat as preview rumor until OpenAI posts. X News
-
@natolambert on inference startups — capital window to own models/supply/post-train before scalps arrive; ties to this week’s open MoE flood. Post
Go deeper
- Full Anthropic memory post + Pokémon / Stainless / LAB figures: components of agent memory
- BioBasicsBench qualitative failure gallery (controls, contrast, colony counts): OpenAI
- DeepMind cryptanalysis PDFs linked from the Pushmeet article thread
- xAI watermark opt-out + audit log path in Console: docs