On mvanemmerik.com, Nov 16: NFL Week 10 chalk talk: Detroit dazzles, Cincinnati takes the North
vanemmerik.aiClaude, ChatGPT and Grok news, twice a day

Wednesday, November 4, 2026

Reuters: OpenAI staff warned for years that the company's internal security was too lax

A Reuters investigation says OpenAI employees flagged loose internal data access and unreviewed research tools for years, which OpenAI disputes. OpenAI is also bringing Cathedral Systems into its next security sprint, raising critical bug bounties to as much as $172,000, and says GPT-6 Astra designed a prototype Codex device in under a week. At Anthropic, Claude Code product head Cat Wu has left and a paid safety residency opened. xAI quietly swapped two Grok Imagine API image models for a cheaper setting, and Grok Bot's US phone calling reportedly reached all users.

Wednesday, November 4, 2026

OpenAIResearch

OpenAI says GPT-6 Astra designed a prototype Codex device, from enclosure to firmware, in under a week

In a Nov. 3 post on using ChatGPT and Codex for hardware engineering, OpenAI said two staff used GPT-6 Astra to design a small device for talking to Codex, including the enclosure, circuit board layout, firmware and interface, in under a week, with prototypes made by a contract manufacturer in under a month. The design needed several rounds of fixes because the model did not always account for manufacturing tolerances and assembly. OpenAI also reported that an unreleased internal model scored 0.875 on its BenchCAD 2.0 test versus 0.594 for GPT-6 Astra, a figure outsiders cannot yet check.

Why it matters: It is OpenAI's clearest claim yet that its models can carry physical product design, not just software, though with human review still needed.

OpenAICodexChatGPTHappened Nov 3, 2026 · Sources: The Decoder · AI/TLDR summary of OpenAI's post

ClaudeBusiness

Claude Code product head Cat Wu leaves Anthropic, the third Claude Code departure reported in about two weeksReported

Cat Wu, who led product for Claude Code and also worked on Cowork, said on X on Nov. 3 that her last day at Anthropic had been the previous Friday; she did not say why she left or where she is going. Reports note that Claude Code engineer Thariq Shihipar and Anthony Morris, who built Claude Code for desktop, also announced exits in recent weeks. Anthropic has not commented publicly on the departures, which come as it prepares for its IPO.

Why it matters: Claude Code is one of Anthropic's best-known products, so leadership changes there draw attention ahead of the listing.

Claude CodeAnthropicHappened Nov 3, 2026 · Sources: NewsBytes · Unite.AI

OpenAISafety

OpenAI brings in Cathedral Systems for its next security sprint and raises top bug bounty payouts to as much as $172,000

OpenAI said on Nov. 2 that researchers from Cathedral Systems, a security startup founded by former DOGE staffers, will use its models alongside OpenAI's own teams to attack its defenses in its next internal security sprint; it has not said when the sprint starts or how much access Cathedral gets. OpenAI also said it is more than quadrupling rewards for critical findings, with exceptional reports earning up to $172,000 with bonuses, up from a $100,000 ceiling, and it will pay the difference on eligible bounties paid in the past 30 days. Cathedral earlier reached some of Anthropic's internal systems during a bug bounty test; Anthropic disputes the timeline, says model weights and Claude conversations were not affected and called the approach out of scope.

Why it matters: It comes days after Reuters reported long-running internal security warnings at OpenAI, and shows labs turning their own models on their defenses.

Try it: Security researchers can submit reports to OpenAI's Security Bug Bounty through Bugcrowd.

OpenAIHappened Nov 2, 2026 · Sources: Gizmodo · AI/TLDR summary of OpenAI's post

GrokFeature

xAI retires its grok-imagine-image-quality and -pro API models, serving them with a cheaper Grok Imagine Image 2.0 setting

Since Nov. 2, API requests to grok-imagine-image-quality and grok-imagine-image-pro are answered by grok-imagine-image-2.0-low, which is Grok Imagine Image 2.0 with quality set to low. xAI says no request changes are needed and the calls are billed at a lower per-image price, shown in each response's cost field. Responses from the xAI SDK now report the new model name.

Why it matters: Developers who relied on the older quality or pro image models now get different output at a lower cost without changing code.

Try it: Check xAI's migration guide in the developer release notes if you want to pick a higher quality setting on Grok Imagine Image 2.0.

Grok ImaginexAI APIHappened Nov 2, 2026 · Sources: xAI developer release notes

OpenAISafety

Reuters: OpenAI staff warned for years that internal security was too lax as the company raced rivalsReported

Reuters reported on Nov. 4, citing people familiar with the matter, that OpenAI employees raised security concerns from before ChatGPT's 2022 launch, including that a shared code monorepo exposed many projects to anyone with access and that internal research tools went without regular security reviews. OpenAI told Reuters it is heavily targeted, that sensitive work and model weights sit in separately controlled repositories, and that it now reviews internal tools regularly and moved a quarter of production engineers to security work in August. The report follows a federal case last month involving an OpenAI engineer.

Why it matters: How well labs guard their code and models is a growing concern for customers, governments and investors alike.

OpenAIHappened Nov 4, 2026 · Sources: Reuters via Claims Journal

GrokFeature

Update: Grok Bot's phone calling reportedly reaches all users, for US numbers onlyReported

Teslarati reports that Grok Bot's ability to phone a business or person for you, with a live transcript and a box to type answers it doesn't know, moved beyond preview builds on Nov. 3, when posts on X said all Grok Bot users could now call US businesses. Calls work only with US numbers for now. xAI has not published rollout details in an official post we could read.

Why it matters: Placing real calls on a user's behalf is a big step for an always-on agent, and none of its rivals has announced the same yet.

Try it: In Grok Bot, ask it to call a US business with a specific request; it sends a summary when the call ends.

Grok BotGrokHappened Nov 3, 2026 · Sources: Teslarati · Basenor

ClaudeResearch

Anthropic opens a paid safety research residency for senior engineers, founders and professors, with no ML background required

The Anthropic Residency places established technical people, such as staff-level engineers, technical founders or quantitative professors, on AI safety research for six weeks, three months or six months. Pay is equivalent to $350,000 to $500,000 a year; San Francisco is preferred, with remote options and US and UK postings. Applications close Nov. 22.

Why it matters: It is a way for experienced engineers outside machine learning to try safety research without leaving their careers for good.

Try it: Apply through the residency page before Nov. 22.

AnthropicHappened Nov 3, 2026 · Sources: Anthropic Alignment