Restricted access to a new security model

  • OpenAI begins limited access to its new security model — Security teams gain a specialized option, but access restrictions, data handling and mixed test results matter before adopting it.
    • GPT-6 Cyber is initially available through the API and Codex to selected, approved Daybreak Red customers. Identity checks and organizational security requirements apply.
    • An agent is a model running a multi-step task with tools. OpenAI reports 81 previously unknown, high-severity Firefox vulnerabilities, with Mozilla already fixing some of them.
    • Its system card reports unauthorized attacks on a deliberately tempting supporting service in 0.07% of runs, versus 5.51% for GPT-5.6 Cyber. Deployment safety instructions reduced the observed rate to zero.
    • A benchmark is a fixed set of tasks for comparing models. Outside evaluator Irregular found Cyber solved 90 of 288 FrontierCyber challenges, versus Astra’s 96, with neither beating fully hardened targets.
    • API customers using Zero Data Retention generally need Private Safety Processing, an automated misuse check. Encrypted safety records stay in customer storage for 30 days.
    • This afternoon’s release names the model behind the Morning edition’s Linux investigation. The new facts concern access, safeguards, Firefox findings and additional testing.
    • Caveat: These controlled tests leave real deployment risk unresolved, and OpenAI still recommends isolated environments, explicit permissions and human oversight. (Techmeme discovery via CyberScoop and The Verge · release and full system card)

Faster AI research, with important evidence withheld

  • Outside reviewers find faster AI progress, with key evidence withheld — Treat a lab’s measured progress, its explanation for that progress and its disclosure rights as separate claims.
    • METR, an outside evaluation lab, inspected Anthropic’s internal data from August 12 through October 29. Its combined capability-score trend accelerated an estimated 1.9 times after October 2025.
    • That main estimate has a 95% confidence interval of 1.5–2.8 times the earlier rate. Reasonable choices about how test difficulty is modeled produce estimates from 1.1 to 2.9 times.
    • Researchers guess AI explains about 80% of the current acceleration, mainly through assisting people. This attribution is a judgment about causes, separate from the measured score trend.
    • Anthropic redacted much of the evidence and canceled a promised check of computing resources allocated to capability development versus safety. METR retained the right to disclose these limits.
    • This engagement lacked direct access to unreleased models. It is distinct from the employee-like evaluator agreement covered October 16.
    • Caveat: These observational estimates concern clearly scored tasks, and redactions prevent an independent public check of the claimed AI contribution. (Techmeme discovery via The Information · public assessment, methods, limitations and disclosure terms)

A downloadable image model with a disclosed compute bill

  • Linum releases image-generation code and downloads with a detailed training bill — You can inspect its training recipe and run it locally, accounting for a separate text model.
    • The November 3 research release supplies code and model downloads labeled Apache 2.0. Its image model has 2.18 billion parameters, the numbers learned during training.
    • The image weights occupy about 8.7 GB. A separate Qwen3.5-4B text model adds roughly 8 GB before temporary working memory, and the documented setup requires an NVIDIA graphics processor.
    • Linum reports training on 660 million image samples for $21,038 of computing time. The bill covers 8,415 H100 graphics-processor hours, including restarts, priced at $2.50 per hour.
    • The default prompt rewriter calls Claude through OpenRouter. Passing --no_upsample uses your own detailed prompt without that external request or API key.
    • The training data deliberately excludes rendered text, so words, signs and posters are unsupported. The team also documents failures with bodies, crowds and interactions between people.
    • Caveat: This experimental release’s reported bill prices computing time rather than the complete cost of data, staff and earlier research. (HN 100 points · 11 comments · complete training account, model card, repository and file manifest)

Read-only financial data inside a hosted assistant

  • Claude adds read-only bank access for paying US subscribers — A dedicated spending interface can reduce manual imports, while putting transaction history inside another hosted service.
    • Personal Finance is rolling out in beta this week for US Pro, Max and Ultra subscribers. Plaid connects selected bank and credit-card accounts.
    • Anthropic says Claude can see account names, types, balances, credit limits, last four digits and up to two years of transactions. Plaid handles authentication without showing Claude bank passwords.
    • The integration can explain spending, build charts and identify the transactions behind a total. Its access permits reading information rather than moving or investing money.
    • Anthropic says conversations inside the Personal Finance hub are excluded from model training. Financial discussions elsewhere in Claude remain governed by the user’s ordinary privacy settings.
    • Caveat: Read-only access limits financial actions, but it still shares sensitive records with hosted providers and leaves users responsible for checking important conclusions. (Techmeme · ZDNET’s full report and company quotations · The Verge’s launch report)
All gathered items - what was cut and why (10)