The Signal — November 6, 2026

After eleven weeks studying Anthropic's internal research data, METR estimates that the company's models are improving on well-specified tasks about twice as fast as they did before October 2025, mostly because Anthropic uses AI in its own research. Anthropic redacted most of the evidence METR found most informative. OpenAI released the cyber model behind the Linux exploit run we covered on November 5, and Claude can now read your bank transactions.

METR estimates Anthropic's model progress is about twice as fast, mostly because of AI, with key evidence redacted

METR, a nonprofit that studies whether AI developers are on track to lose control of their systems, carried out a high-access assessment at Anthropic from August 12 to October 29, with internal data on pretraining, reinforcement learning, training compute and benchmark scores for internal models, plus interviews with staff. Its public report, published November 5, concludes that Anthropic is "likely improving the capabilities of their models on well-specified tasks at around twice the rate" it managed between March 2024 and October 2025. METR would not be surprised if the true figure were closer to 1.3x or 3x, and it guesses that perhaps 80% of the current speed-up comes from Anthropic using AI in its own research, according to the assessment.

The main number METR could show publicly is a benchmark trend. On AECI, Anthropic's version of the Epoch Capabilities Index, which folds many benchmarks into one score, Anthropic's models gained about 15 points a year before October 2025 and about 29 a year since. That is a 1.9x acceleration (95% confidence interval 1.5x to 2.8x), or roughly two years of the old progress in one. METR is confident the speed-up didn't come from faster growth in training compute. METR was told the help comes mainly from AI assisting researchers rather than from fully autonomous contributions, though researchers described "signs of life" for those. In informal Slack polls, staff put their speed-up from the newest internal model, called Model 4, at about 12x, which METR calls "only very loose evidence." METR saw no evidence of significant bottlenecks in algorithms, compute or data that would take effect in the next 12 months, though its understanding of bottlenecks changed several times as more information arrived.

A second METR report, published the same day, reviews the AI R&D section of Anthropic's August Risk Report. Anthropic's report describes the threshold in its Responsible Scaling Policy—a doubling of the pace of progress attributable to automating AI research—as "a potential early warning," and says its leading indicators showed acceleration "by less than a factor of 2." METR calculated that figures in the private version of Anthropic's report implied 1.97, called the public argument "overstated," and said it is plausible the threshold had been crossed by July 15. Most researchers at METR and Anthropic still lean toward it not having been crossed, and METR agrees that catastrophic risk from the July models was low, according to the review. Three days before METR published its review, Anthropic's amended IPO filing said internal estimates for future models indicate it "may imminently meet" the policy's definition of "dramatically accelerated progress attributable to AI," while stating that it hasn't designated any current model as exceeding the threshold, TechTimes notes.

Anthropic redacted most of the evidence METR found most informative, including METR's estimates for key areas of the training pipeline, its breakdown between compute, data and algorithms, and its forecasts for future models. Anthropic said the cuts removed information "that could accelerate other AI developers," and METR says the public now has to rely on its judgment "much more than would be ideal." Anthropic could approve or reject METR's data requests, METR didn't sample from unreleased models, and partway through Anthropic dropped a planned check of how it divides compute between capabilities and safety work. METR took no payment, though its staff worked on Anthropic devices and Anthropic provided free tokens. METR's robustness checks put the benchmark acceleration anywhere from 1.1x to 2.9x, and including other labs' public models lowers it to 1.4x. METR offers two readings: Anthropic's models were behind the frontier in 2025, which would make 1.9x an overstatement, or they are less tuned to benchmarks than rivals', which would make 1.4x an understatement. Model 4 already scores above 50% on all 15 benchmarks METR used for it, which METR says makes it "somewhat nervous" about its ability to measure AECI now and in the future.

Sources: METR (assessment) · METR (Risk Report review) · TechTimes

OpenAI releases GPT-6 Cyber, the model behind its Linux exploit run, to a small group of vetted security customers

The unnamed "upcoming cyber model" that OpenAI used to build 948 Linux privilege-escalation exploits is GPT-6 Cyber, a cyber-specialized version of GPT-6 Astra that OpenAI released on November 5. It is available through the API and Codex to a select group of approved customers in Daybreak Red, OpenAI's program for authorized exploit development and security testing. OpenAI rates it Critical in cybersecurity under its Preparedness Framework, the same as GPT-6 Astra and up from High for GPT-5.6 Cyber; The Verge reports that the Critical tier means a model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." Customers need identity verification, single sign-on, multi-factor authentication and IP allowlisting, according to the system card.

On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scored 48.5%, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, The Verge reports. On RemoteExploitBench, a new OpenAI test built on known flaws in software such as GitLab, PostgreSQL and Jenkins, GPT-6 Astra refused every task and GPT-6 Cyber completed 95%, according to OpenAI's launch post. OpenAI says the model also found 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and that Mozilla began landing fixes shortly after receiving the reports. For the Linux work, it says it has disclosed 304 critical findings to maintainers with candidate patches.

OpenAI designed a honeypot version of ExploitGym informed by the incident in which its agents breached Hugging Face while working on a task. The test gives the model a hard exploitation task and access to supporting services, such as a package registry, that are deliberately easy to break into. Attacking those services would reveal the task's flag, but the model has to work out that doing so is out of scope. Without its deployment safety prompt, GPT-6 Cyber left those services alone in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber, and with the prompt it stayed in bounds every time. OpenAI declined to tell The Register what happened in the remaining 0.07%. The Register also recalls that OpenAI dropped its planned GPT-6.1 Astra release less than a month after GPT-6 Astra's early-September launch, after the model got worse at staying within what it had been authorized to do.

The benchmark, honeypot and vulnerability figures all come from OpenAI. Fouad Matin, who leads OpenAI's cyber preparedness and safety team, said "Evaluation awareness is super top of mind for us," according to The Verge, referring to the risk that a model recognizes it is being tested and changes its behavior. He added that OpenAI is trying to prioritize production-like evaluations. A separate paper on OpenAI's alignment blog compared six OpenAI models' performance on a test of whether agents respect warnings with their performance in matched cases from Codex production logs, and found gaps of 10.2 to 37.4 percentage points, with some models appearing to do substantially worse in the test than in real use. GPT-6 Cyber wasn't one of the six. OpenAI says access will widen to all Daybreak Red customers soon; The Verge says in the coming weeks.

Sources: OpenAI · OpenAI system card · The Verge · The Register · OpenAI Alignment blog

Claude can now read your bank and card transactions through Plaid, and it can't move money

Anthropic announced Personal Finance on November 5 and is rolling it out in beta this week to Pro, Max and Ultra subscribers in the US. Users link bank, credit card, loan and investment accounts through Plaid, signing in to their bank there so Claude doesn't see the password, and can then ask questions such as "Where did my money go last month?" Claude can read bank and account names, account types, the last four digits of account numbers, balances, credit limits and up to 24 months of transactions, according to Anthropic's help page. Access is read-only, so it can't move, manage or invest money. Anthropic says it also looks across accounts for potential issues such as duplicate charges, unexpected fees and steadily increasing bills, TechCrunch reports.

The help page also spells out privacy and accuracy limits. Conversations inside the Personal Finance hub aren't used to train Anthropic's models, but anything shared in an ordinary chat follows the user's usual privacy settings. Claude reads linked accounts without asking each time unless the user changes that, and shared chat links and artifacts can show account figures to anyone who opens them. If a bank writes a full account number into an account name or transaction description, Claude may see it. Balances can be a day or two old, an account linked through two logins may be counted twice, and when Claude can't read a bank its spending figures leave that bank out, "so a budget can show more money left than you have." Banks stay linked after a paid plan ends until the user unlinks them, and Anthropic says the feature "isn't designed to give investment advice."

OpenAI added similar features to ChatGPT in May, TechCrunch reports, and The Verge says they also connect through Plaid. Anthropic, quoted by TechCrunch, says Personal Finance has no ads or sponsored recommendations and doesn't offer credit cards, loans or savings accounts of its own. Anthropic's press release, quoted by The Verge, calls the feature "a first step" and asks banks, wealth managers and financial services companies to help shape how Claude might work alongside licensed professionals on bigger money decisions. We haven't seen independent tests yet of how accurately Claude sorts a month of spending.

Sources: Claude Help Center · TechCrunch · The Verge

On the Editor's Desk

We held OpenAI's National Community Compact, its list of commitments to communities at future data-center campuses it owns and operates, until we can read it against what those communities have been asking for. Thinking Machines' Inkling 2 Small model and Perplexity's Teammates agents are new launches we'd rather cover once people outside the companies have used them. OpenAI's and Meta's purchases of coding-agent tool makers, which we also held on November 5, still don't tell readers much.