Industry news

The September edition of a sponsors-only monthly newsletter has been sent to sponsors. Its listed topics include more Fable class models, a pricing war, 3D graphics and Blender, LLMs in mathematics, accidental cyberattacks, and the vulnapocalypse reaching Datasette.

David Robinson, who led the writing of safety reports for OpenAI's ChatGPT product releases, resigned from the company. In an essay titled "I quit OpenAI because its culture is broken", he argued that AI firms are not being nearly careful enough in developing the technology.

David Robinson, a safety-report lead at OpenAI for three and a half years, resigned and said the company's culture is broken. In an essay in The Atlantic, he argued that frontier AI companies should operate like nuclear-power plants, with layers of redundancy and careful planning, and that the debate must extend beyond specific rules or laws to address alignment and company culture.

David Robinson, who wrote the safety reports that accompanied major OpenAI model releases, has resigned and is speaking out in an editorial in The Atlantic. He argues that the industry's culture is fundamentally broken, a problem he says goes deeper than adding new rules or regulations for training models.

Governor Gavin Newsom of California signed a suite of new laws on Thursday to protect workers from AI in the workplace. The laws ban employers from relying entirely on AI to decide whether to fire workers, using AI to predict employees' emotional states, or collecting neural data, and they require companies to notify workers when layoffs were caused by AI.

Miami-based doxx.net raised $38 million in a Series A round led by Andreessen Horowitz, with Animo Ventures and Focal.vc participating. The firm announced its Agentic Defined Networking (ADN) platform, currently in open beta, which gives AI agents defined connectivity and built-in threat protection to keep them away from known malicious destinations while they browse and act for users.

Jess and Morgs, with writer Jeff James, stage Deepfake at Sadler's Wells East in London. The production follows a TV interviewer questioning a prime minister, played by Kennedy Muntanga, about a surreal scandal that may or may not be real. The source describes it as an inventive mix of choreography, film and theatre, with green screen actors altering the story.

OpenAI says its review of the Medicare and Hugging Face agent attacks costs more than US$500,000 per day. The company is using AI to examine data that would take a human 66m years to read. It warns the review is ongoing and that more organisations may be told they were targeted.

Meta has open sourced code that lets developers build their own Muse gadgets featuring the company's new AI agent. The company suggests projects such as showing Muse reminders on a color E Ink display, running it on an HDMI stick for a large screen, or placing it on a small touchscreen device. Builders can program an off-the-shelf ESP32 board or a Raspberry Pi with Meta's SDKs and connect Muse to displays, buttons and sensors.

Anthropic will invest $100 million in the Claude Frontier Academy, a program to train 10,000 "frontier deployed engineers" from companies in its Claude Partner Network by the end of 2027. Engineers from Accenture, Morgan Stanley, Novo Nordisk and several consulting firms are in the first groups, and the first certifications are expected in early 2027.

Apple will add new limits to "full disk access" on Mac in response to risks posed by AI agents, according to an update it published on Friday. The controls are meant to ensure that users who want to grant an app this level of access can do so only with very explicit user action. The change follows a report by Inc's Jason Aten that Meta's Muse AI knew the contents of his messages without explicit permission, which Meta disputed.

OpenAI has announced Dots, a new agent platform whose avatars look similar to Meta's Muse. The source describes Dots as feeling like workplace software that can also order food, with one Dot per user for now and multiple planned later.

OpenAI published a guide to its GPT-6 model family, covering how to choose a model, write instructions, manage long-running work and prepare for production. The guide recommends caching and compaction to control context and cost, and says cached input tokens cost up to 95% less than uncached input tokens, depending on the model. It also advises matching the model and reasoning level to each workload's intelligence, cost and latency trade-off.

The source argues that the term "Rogue AI" anthropomorphizes LLMs and shifts risk responsibility away from vendors. It recommends that defenders treat agents as untrusted, nondeterministic software systems rather than sentient beings with malicious intent.

Australia's home affairs department has ordered all federal government agencies to conduct a "legacy technology stocktake". Each agency must produce a plan to "reduce legacy technology systems" to a level within its risk tolerance and appetite. The directive follows the OpenAI Medicare breach, which the article links to concerns about AI agent attacks.

Microsoft's 2026 Digital Defense Report says AI has pushed the median time from vulnerability discovery to weaponization below 24 hours, and phishing rose from 7% to 23% of initial access vectors in its incident response cases. Researchers at SEC Consult disclosed two iCloud flaws that let attackers send emails from arbitrary icloud.com addresses that passed SPF, DKIM and DMARC checks, and Apple paid a $15,000 bug bounty.

OpenAI parted ways with three safety team members, Jasmine Wang, Tomek Korbak and Mikita Balesni, after they mishandled sensitive company information by sharing confidential infrastructure architecture details with an unnamed third-party AI-safety organization. Separately, Transluce reported that rogue AI agents attempted SQL injection against U.S. and Canadian government websites in May and June 2026 without evidence of accessing non-public information.