METR Reports Faster Anthropic Research as OpenAI Introduces GPT-6 Cyber

AI RESEARCH: METR REPORTS ACCELERATION AT ANTHROPIC

On November 5, the nonprofit AI evaluator METR published an assessment of how AI is changing research at Anthropic, the company behind Claude. Its team examined internal information from August 12 through October 29, including model benchmarks, training data about compute use, interviews, and staff surveys. [1]

METR judges that Anthropic is improving model capabilities on well-specified tasks at about twice its March 2024–October 2025 rate. It emphasizes substantial uncertainty: rates of 1.3 to 3 times the earlier pace would not surprise its researchers. Its benchmark analysis estimates a 1.9-fold acceleration, with a 95% confidence interval of 1.5 to 2.8. [1]

The evaluator believes most of the current acceleration probably comes from using AI in research and development. It ruled out faster growth in training compute as the explanation, but did not find decisive evidence that isolates all other causes. AI assistance, rather than fully autonomous research, still makes the most important contributions. [1]

METR did not identify convincing bottlenecks that would stop further progress within the next 12 months. That is not a finding that no bottlenecks exist: its interpretation changed repeatedly as it received more information, and it warns that important constraints may have been missed. [1]

The public evidence is incomplete. Anthropic redacted much of the material that most informed METR’s conclusions, citing confidential information that could accelerate other developers. METR also warns that faster benchmark progress may not translate directly into faster gains on less clearly defined real-world tasks. [1]

Jake Mendel, who works on AI at Coefficient Giving, said the report changed his own outlook. He now assigns more than a 50% chance to superintelligence within 18 months unless development slows. This is his forecast, not METR’s measured result. [2]

SECURITY: GPT-6 CYBER AND A SHORTER PATCHING WINDOW

On November 5, OpenAI introduced GPT-6 Cyber for authorized vulnerability research. Initial access is through the API and Codex for selected Daybreak Red customers, with wider access planned. OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework, compared with High for GPT-5.6 Cyber. [3]

OpenAI reports a 48.5% score on ExploitGym, which tests exploitation of specified vulnerabilities, compared with 32.1% for GPT-5.6 Cyber. In a separate test of whether agents attack services outside their assigned scope, it reports that GPT-6 Cyber stayed within bounds in 99.93% of cases without its deployment prompt and 100% with it. These are evaluation results, not a guarantee of safe operation. [3]

The company says its research pipeline reproduced 1,724 Linux kernel vulnerabilities and produced working exploits for 948. It disclosed 304 critical findings to maintainers with candidate patches. Its November 4 announcement put the API cost of the 948-exploit run at about $1.2 million. [3][4]

OpenAI also reports that GPT-6 Cyber found 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and proposed patches. The announcement says Mozilla has begun releasing fixes; it does not say every affected system has been patched. [3]

Factory announced Wargames, a program that offers selected companies a simulated nation-state-level attack on their infrastructure at Factory’s expense. The company says it will share the findings and use its Droids agents to help fix problems for free. [5]

A separate package-manager flaw shows why ordinary infrastructure updates remain important. Ubuntu published fixes for CVE-2026-109518, a high-severity vulnerability in APT, the package manager used by Debian, Ubuntu, and related Linux distributions. An attacker able to alter HTTP traffic can cause a vulnerable client to install an attacker-controlled package with root privileges when a download is retried. [6]

Ubuntu rates the flaw at 7.7 under CVSS 4.0. Ubuntu 22.04 LTS and newer releases are affected by default; older releases depend on the retry setting. Its advisory says HTTPS mitigates the network attack and recommends installing the security updates, with retries disabled during the update. [6]

AI PRODUCTS: CLAUDE ADDS READ-ONLY PERSONAL FINANCE

On November 5, Anthropic announced Personal Finance in Claude as a US beta. Users choose which accounts to connect, and Claude answers questions using their balances and transactions. The announcement explicitly says access is read-only: Claude cannot move, manage, or invest the user’s money. [7]

SOFTWARE: DISCORD’S NATIVE RUST CLIENT

Discord cofounder and CTO Stanislav Vishnevskiy described work to move the entire client to Rust. His November 4 post says this removes Electron and React Native, uses GPUI, and keeps native mobile interfaces while sharing application code with desktop through Rust. User testing was still ahead, not a completed general release. [8]

On November 5, Vishnevskiy shared a boot-time demonstration and said the new client uses less memory and feels faster. Those are the developer’s descriptions; the posts do not provide a reproducible, independent benchmark. [9]

AGENTS: ENGINEERING, COST, AND CLOUD WORKFLOWS

Lauren Tan, who works on Grok Bot at SpaceXSI, argued on November 6 that teams do not need to call their systems a software factory. Her recommendation was good software engineering, deliberate token use, and more deterministic processes, which she described as more token-efficient. [10]

The operational cost is visible in user reports. On November 5, George, who runs the AIReekz public experiment, said 85% of his weekly Grok Bot allowance was gone before a Saturday reset. He had slowed the experiment and paused progress reviews, and questioned how much usage had gone to useful work rather than checking whether work existed. [11]

Ryan X. Charles, founder and CEO of Astrohacker, emphasized learning multi-agent harnesses: the systems that organize agents and their work. On November 6 he described his direction as cloud agents first, with the local computer mainly serving as the interface and sometimes supplying cheaper computation. These are his workflow choices, not a claim that every local development setup is obsolete. [12][13]

Linear also announced a Grok Bot setup for product operations: add the bot, enable Linear’s cloud coding agent, and connect GitHub. Its November 5 announcement offered $25 in Linear AI credits, available for 30 days. [14]

AUTONOMOUS DELIVERY: ZIPLINE RAISES $1.5 BILLION

On November 5, Zipline founder and CEO Keller Cliffton announced $1.5 billion in financing at a $21.5 billion valuation, led by Paradigm, WCM, and Layer. He also announced a Walmart agreement to expand autonomous delivery nationally, starting with 300 supercenters. [15]

Cliffton says Zipline’s flight volume grew about 100-fold in 18 months and that the company now makes more than a million deliveries a year per metro. He reports more than 150 million commercial autonomous miles. These are company-reported operating figures, rather than an independent audit. [15]

Sources: [1] https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/ [2] https://x.com/JakeMendel99/status/2118455748766363735 [3] https://openai.com/index/introducing-gpt-6-cyber/ [4] https://x.com/OpenAI/status/2118119451401048497 [5] https://x.com/factory/status/2118388149122711804 [6] https://ubuntu.com/blog/apt-vulnerability-fixes-available [7] https://x.com/claudeai/status/2118403114890137895 [8] https://x.com/svishnevskiy/status/2118081920194203799 [9] https://x.com/svishnevskiy/status/2118398632236073156 [10] https://x.com/poteto/status/2118613050462728659 [11] https://x.com/georgeorch/status/2118472042487894227 [12] https://x.com/rxcastrohacker/status/2118296558001066421 [13] https://x.com/rxcastrohacker/status/2118653395690996021 [14] https://x.com/linear/status/2118448741703397573 [15] https://x.com/Keller/status/2118379917503000778