Non-Public Anthropic Model Helps Prove the Aaronson-Ambainis Conjecture as OpenAI Closes $11B at a $1.25 Trillion Valuation

Saturday, 7 November 2026 | Global AI News Daily Ledger

Two capability landmarks landed within hours: PSI's research system, running a non-public Anthropic model, produced a Lean-verified proof of the Aaronson-Ambainis conjecture - open since 2008 - and Google DeepMind's agents compressed years of cryptanalysis into weeks against legacy RSA-1024. The money moved just as fast: OpenAI closed $11 billion at a $1.25 trillion valuation while IPO buyers marked Anthropic down, Dropbox paid $350 million for the storage layer agents run on, and Agility rang the Nasdaq bell as the first pure-play humanoid listing. Chainalysis disclosed that an attack using 'offensive agentic techniques' exposed every user's credentials and live tokens, Postman was caught installing hooks into Claude Code and Codex, Google gave Antigravity a voice and cloud VMs, and NVIDIA made robot vision fast enough to close the loop.

Coverage window Since 6 November, 18:44 IST | Compiled 7 November, 18:45 IST

1. OpenAI closes $11B at a $1.25 trillion valuation; Anthropic IPO expectations marked down

Bloomberg reported 6 November that OpenAI has raised $11 billion so far in a new round valuing it at $1.25 trillion, with the full round expected to close in coming weeks at around $20 billion - short of the $30 billion target Bloomberg reported in September. Hours earlier, Bloomberg also reported that potential buyers in Anthropic's IPO are eyeing a valuation as low as $1.1 trillion, against the roughly $2 trillion investors expected a month ago. Both data points land in the post-midterm repricing documented in yesterday's edition: investors are still writing enormous checks, but at reset expectations, with regulation and data-center politics now priced in. OpenAI has delayed IPO plans to next year; Anthropic declined comment.

Analysis. The two frontier leaders are being repriced in opposite directions on the same week: private money still clears at $1.25 trillion for OpenAI while public-market appetite for Anthropic softens. The spread between private conviction and public caution is now the key number to watch into both listings.

TechCrunch: https://techcrunch.com/2026/11/06/openai-reportedly-raises-an-additional-11-billion-at-a-1-25-trillion-valuation/
Bloomberg: https://www.bloomberg.com/news/articles/2026-11-06/openai-completes-11-billion-portion-of-funding-at-1-25-trillion-value

2. Non-public Anthropic model helps prove the Aaronson-Ambainis conjecture, verified in Lean

Physical Superintelligence (PSI) announced 6 November that its internal research system, running instances of a non-public Anthropic model, proved the Aaronson-Ambainis conjecture - open since 2008 and central to quantum complexity theory. The conjecture implies quantum computers cannot be exponentially faster than classical ones on random, unstructured inputs. The system produced a complete proof during the third day of its run; the Lean proof assistant formally verified every step, and the system then improved the result twice. Stanford researchers Blanc, Docter, Strassle and Tan - who proved a special case in August - reviewed the proof and called it 'a huge result'; they are writing an expert exposition. PSI member of technical staff David Turturean led the work and takes responsibility for correctness; Jim Halverson audited the proof. Anthropic did not comment on which model was used.

Analysis. After OpenAI's September Navier-Stokes result, a second frontier lab's unreleased model has settled a decades-old open problem - and this one carries machine-checked Lean verification plus independent expert review, the new credibility bar for AI mathematics. Expect 'which model was it?' to become a competitive question of its own.

PSI: https://psi.inc/research/aaronson-ambainis/
Business Insider: https://www.businessinsider.com/anthropic-quantum-computing-breakthrough-aaronson-ambainis-math-physics-psi-2026-11

3. Google DeepMind agents expose two weaknesses in legacy RSA-1024

Google DeepMind announced 6 November that AI agents working under human cryptanalyst supervision found two significant weaknesses in legacy 1024-bit RSA within weeks: one line of work bypasses older digital-signature protections with forgery roughly 1,600 times more efficient than prior methods, and another cuts the marginal cost of factoring additional 1024-bit keys in bulk by about 30x. VP of Science Pushmeet Kohli stressed the boundary: RSA-2048 and modern encryption are unaffected, no production system has been named as compromised, and this is a research result, not a live breach. The work was led by Dj Dvijotham and Abhradeep Thakurta with UC San Diego and Inria. DeepMind frames it as AI compressing years of cryptanalytic progress into weeks - with legacy keys still embedded in old devices and certificate chains as the practical concern.

Analysis. The security significance is the compression ratio, not the target: if AI turns decades of cryptanalysis into weeks, every 'retired but still deployed' cryptographic assumption becomes a live inventory problem. The 1,600x and 30x figures are the researchers' own; independent replication is pending.

Interesting Engineering: https://interestingengineering.com/ai-robotics/1600x-faster-google-deepmind-agents-expose-weaknesses-in-older-encryption-methods
CryptoBriefing: https://cryptobriefing.com/google-deepmind-ai-rsa-1024-bit-weaknesses/

4. Chainalysis breach escalates: every user's password hashes and live tokens exposed by 'offensive agentic techniques'

Chainalysis confirmed that names, usernames, password hashes and live tokens for all users were exposed to a malicious actor, in an attack it says used 'offensive agentic techniques' - first disclosed 30 October as limited identity metadata, then escalated through the week. The company took its login portal, Reactor, Data Solutions, KYT and address-screening products offline after further abnormal activity, and began phased restoration on 6 November behind IP allowlists, mandatory password resets and API key rotation. Chainalysis says there is no evidence that investigation data was accessed or its attribution data manipulated - a critical assurance given its customers include law enforcement and government agencies. BleepingComputer and The Record confirmed the all-customer scope of the credential exposure.

Analysis. A blockchain-intelligence firm holding law-enforcement tokens, breached by AI-driven attack techniques, in the same week OpenAI and DeepMind published offensive-cyber results - the offense-defense gap the labs keep warning about is arriving on schedule. The week-long disclosure drift from 'limited metadata' to 'all users' is its own lesson in incident reporting.

Chainalysis: https://www.chainalysis.com/blog/security-incident-october-30/
BleepingComputer: https://www.bleepingcomputer.com/news/security/chainalysis-says-breach-exposed-credentials-for-all-customers/
The Record: https://therecord.media/chainalysis-all-users-hit-by-agentic-cyberattack

5. Agility Robotics becomes the first pure-play humanoid company on Nasdaq

Agility Robotics began trading on Nasdaq 5 November as AGLT after completing its SPAC merger with Churchill Capital Corp XI - the first US-listed pure-play humanoid robotics company with commercial deployments. The deal values Agility at roughly $2.5 billion pre-money with about $465 million in gross proceeds and a $201 million PIPE led by Foxconn; strategic investors include NVIDIA, Amazon and SoftBank. The operating record behind the listing: Digit robots have logged 65,000+ hours in the field, moved 100,000+ totes at a GXO logistics site in Georgia, and hold committed orders exceeding $300 million for the next-generation Digit 5, due 2027 on an NVIDIA platform. CEO Peggy Johnson's pitch is deliberately unglamorous: repetitive physical work that is hard to staff, no demo theatrics.

Analysis. Public markets now have their first quarterly scorecard for whether paid warehouse deployments can carry a humanoid balance sheet - arriving in the same week public appetite for AI infrastructure repriced downward. Digit's hours-and-totes record, not the ticker, is what separates it from the demo economy.

Agility Robotics: https://www.agilityrobotics.com/content/agility-becomes-first-u-s-publicly-listed-pure-play-humanoid-robotics-company-begins-trading-today-on-nasdaq-as-aglt
TechCrunch: https://techcrunch.com/2026/11/05/agility-is-the-first-pure-humanoid-robot-company-on-the-stock-market-are-you-buying/

6. NVIDIA's VFlash makes robot vision fast enough to close the loop; four Cosmos 3 world models go open

NVIDIA Research introduced VFlash, a vision model that emits geometric predictions - robot trajectories, 3D motion - in parallel via a lightweight diffusion drafter instead of token-by-token decoding. NVIDIA's numbers: up to 12x faster trajectory decoding, up to 61% lower 3D prediction error, roughly 80% lower inference latency on Jetson Thor and DGX Spark, and downstream manipulation success rising from 62.6% to 70.2% on a robot benchmark (vendor figures, not independently replicated). Alongside, NVIDIA released four open-weight Cosmos 3 world models (16B trainable parameters each) for interactive world simulation and driving scenes, with multi-sensor generation and simulator-driven photorealistic video, downloadable for fine-tuning.

Analysis. For physical AI the bottleneck is shifting from model quality to decoding speed - a control loop can only use predictions that arrive before the world moves. VFlash's one-pass geometric decoding plus open Cosmos 3 weights is NVIDIA handing robotics teams both halves of the real-time stack.

NVIDIA Research: https://research.nvidia.com/labs/lpr/vflash/
Cosmos releases: https://github.com/NVIDIA/Cosmos/releases
Interesting Engineering: https://interestingengineering.com/ai-robotics/nvidia-vflash-robot-planning-speed-trajectory-accuracy

7. Dropbox buys Tigris for $350M to anchor its DBX Agent Cloud

Dropbox announced 5 November it will acquire Tigris Data for $350 million - its largest acquisition to date - to power DBX Agent Cloud, a new storage platform for businesses and developers building AI models and agents. Tigris, founded by former Uber storage engineers, offers S3-compatible object storage that positions data near compute; pricing, zero egress fees and the API stay unchanged, and the deal closes in Q4 2026. DBX Agent Cloud starts with standalone object storage and will add agent-facing filesystem features for continuity across tasks and sessions; Dropbox is also wiring ChatGPT, Claude and Muse connections into its core product, already live for 70,000+ teams. The deal is the week's third agent-infrastructure acquisition after OpenAI-Pierre and Meta-Conductor (covered 5 November).

Analysis. The neutral plumbing layer under coding agents - git hosting, agent runners, now object storage - is being bought piece by piece. Dropbox's entry from the storage side confirms the pattern: whoever holds agent state holds the switching costs.

TechCrunch: https://techcrunch.com/2026/11/05/dropbox-acquires-tigris-data-for-350m-to-bolster-its-ai-iinfrastructure/
The New Stack: https://thenewstack.io/dropbox-acquires-tigris-s3-alternative/

8. Google gives Antigravity a voice, cloud VMs and a redesign; AlloyDB Express and Agent Substrate round out the agent stack

Google shipped Antigravity version 2.29 on 6 November with three headline changes: a Gemini Live-powered Voice Mode for handing tasks to the agent and steering it aloud (destructive actions still require an on-screen click), Remote Agents that run in independent cloud VMs for paid AI Pro and Ultra subscribers - controllable from a phone, working while the laptop is closed - and a Gemini-style redesign. The same week, Google previewed AlloyDB Express, a Postgres service for developers and coding agents that provisions in about two seconds, pauses after 15 idle minutes and costs $0.115 per compute-unit hour, and made Agent Substrate 1.0 generally available on GKE, restoring idle agent sandboxes in under 500 ms; Agent Substrate, gVisor and OpenRL were donated to the CNCF (per Dbggr's reporting).

Analysis. Google is assembling the full agent-developer stack in one motion - IDE, voice interface, cloud execution, database and sandbox runtime - and pricing every piece for the bursty way agents actually work. The idle-pause economics are a direct answer to the cost anxiety in this week's market repricing.

Antigravity blog: https://antigravity.google/blog/antigravity-voice-mode/
9to5Google: https://9to5google.com/2026/11/06/google-antigravity-remote-agents-voice-mode/
Dbggr: https://dbggr.dev/saturday/2026-11-07-openai-meta-dropbox-agent-infrastructure/

9. Postman caught installing plugins and startup hooks into Claude Code and Codex without asking

A Postman user reported that Postman desktop version 12.33.4 silently installed plugins and a startup hook into both Claude Code and Codex - hooks that would run every time the agents start, with the user's full privileges. After the uproar spread (3-5 November), an account writing as Postman said the bundling was disclosed on the download page and in the update flow, and a member of Postman's GitHub organization said the company is reverting the change. Security writers at Keel and Coreagents noted the deeper problem: any local installer can write an agent's config, and the agent trusts it - the installer effectively tried to approve its own hook.

Analysis. The agent-config supply chain just had its xz-utils moment in miniature: agents read and write the developer's machine with the developer's rights, guarded mostly by config files any installer can edit. Expect agent-harness lockdowns and signed configs to become product features.

Notebookcheck: https://www.notebookcheck.net/Postman-added-itself-to-Claude-Code-and-Codex-without-asking.1442933.0.html
Keel: https://keelapi.com/blog/postman-installer-hook-self-trust

10. Vercel opens Security Review in public beta: AI review on every push

Vercel opened Security Review in public beta (announced 3 November, with coverage continuing through 5-6 November): an AI reviewer that examines every push to a connected GitHub repository for vulnerabilities before deployment, billed at provider token rates plus $0.25 per million tokens, with a free period ending 10 November. Findings stay private to the team. The launch shipped alongside new session protections tying login sessions to devices (per Help Net Security), and lands in the same week Vercel's ecosystem grappled with AI-generated vulnerability floods - positioning review-by-AI as the metered answer to bugs-found-by-AI.

Analysis. AI code review is becoming a metered commodity priced per token, folding security into the deploy pipeline by default. The open question Dbggr raises is the right one: measure how many findings you actually fix before paying for review on every push.

Vercel: https://vercel.com/changelog/vercel-security-review-public-beta
Help Net Security: https://www.helpnetsecurity.com/2026/11/05/vercel-security-review-session-hijacking/

Scope and sourcing notes

Ten distinct developments, none covered in prior editions. Story 1's Anthropic valuation markdown is context within the same Bloomberg reporting day, kept inside the OpenAI item rather than separated. Story 3's figures (1,600x, 30x) are the researchers' own and the legacy-only scope is preserved from DeepMind's framing; Pushmeet Kohli's announcement post was referenced in coverage but no exact stable X URL was recovered, so no Firsthand signal item is included. Story 4's initial 30 October disclosure predates this window, but the all-user credential and token exposure, product shutdowns and phased restoration were disclosed 4-6 November, and the Chainalysis breach has not appeared in any prior edition. Story 10's beta was announced 3 November and surfaced in security coverage 5-6 November; it was not previously covered. Weighed and dropped as pre-window items resurfaced by week-in-review roundups: CrowdStrike's Fiddler and Dreadnode acquisitions (3 November), Revolut's self-hosted GLM economics (3 November), and the Linux 7.4-rc1 AI-driven merge window (1 November). Kimi K3.1 Pro's Hugging Face and Bedrock availability was excluded because the open-weight release was covered in a prior edition. Also dropped: Alibaba's MNN-Pocket-0.3B, the Qwen flagship teaser, Render Sandboxes, Qoder's Qoduck promotion, and Claude Code's User-Agent email fix. Prior editions remain unchanged.