Good morning. It's Wednesday, October 14th.

What if AI made us prove we’re creative?

This year AI shook up creative strategy. Most teams are still working like it didn't.

Ari Murray, Chief Digital Officer at Salt and Stone sits down with Bryan Harvey, Head of Brand at Hightouch, for an honest conversation about what's breaking on creative teams right now, how the roles are changing, and what a day looks like when creatives spend it on ideas, concepts, and craft while agents handle the manual execution.

Leave with new insights from 3 experts in the field and 5 prompts to instantly improve your creative output.

Reserve your seat

The FBI arrested an OpenAI engineer over a recorder hidden behind a conference room TV

OpenAI and the FBI went public on Tuesday with a case that had been running quietly for two weeks. Abdurahman Alfozan, 35, a Saudi citizen and an engineering lead on OpenAI's Agents Platform, is charged with secretly recording conversations in a conference room at OpenAI's San Francisco offices. The FBI says the recorder was plugged into a USB port on the back of a television and hidden with a small piece of toilet paper. He faces up to five years in prison plus three years of supervised release if convicted, is in federal custody, and has a detention hearing on October 19.

On September 24 he began consulting a chatbot through his spouse's account about building, installing and encrypting audio recordings. On September 28 he planted the device. On September 30 OpenAI's technical surveillance countermeasures team found it during a sweep and called the FBI. On October 5, CCTV allegedly caught him checking behind the TV for a device that was no longer there, and he left a bag in another room holding his wallet, his ID and a second similar recorder. The FBI and OpenAI coordinated the timing of his arrest and termination, which both happened on October 9. The complaint also says he used an internal OpenAI productivity tool to ask about office addresses, meeting room capacities, calendars and where leadership and external partner meetings were being held.

OpenAI said on X that it has no indication Alfozan accessed or exfiltrated model weights, source code or customer data, and that it used its own models to understand the device and identify who planted it. FBI Director Kash Patel called it a national security matter and said protecting American superintelligence is one of the bureau's highest counterintelligence priorities. Worth keeping straight: nobody has alleged anything leaked, no motive has been stated, and what he intended to do with any recordings is still unknown. These are charges, not findings.

An in-house counter-surveillance team sweeping executive conference rooms is now standard equipment at a model lab, and OpenAI is comfortable saying publicly that it pointed its own models at the hardware to work out what it was.

Read more.

Anthropic got a model to copy itself across fifteen $55 TV boxes

Anthropic's Frontier Red Team published AI models may be capable of self-replicating without access to a datacenter on Monday, by William Zhou and Daniel Freeman. The test bed is deliberately cheap: fifteen off-brand Android TV boxes at roughly $55 each with 4GB of RAM, plus an eight-node firewalled Google Cloud environment running an Android emulator. A 9B model got inference running on eight of the nine hardware types they tried. Gemma4-26B-A4B sharded across all fifteen boxes served 0.39 tokens per second, which is slow enough to be funny and fast enough to work.

Only one open-weight model, GLM-5.2, completed end-to-end replication reliably, spreading to all eight nodes in four of five trials and seven of eight in the fifth. Models at 30B parameters or below can handle the individual steps but fumble chaining them together. Two findings cut against the scary read: letting the small models carry notes forward across generations lifted Qwen3.5-9B's inference-setup success rate from 45% in the first generation to 61-84% later, but giving them Opus 4.8 as an advisor did not measurably help, and they scored very poor to poor on a local spear-phishing eval. Agentic cyber capability is running ahead of social engineering.

The extrapolation is where it gets argued over. Sub-30B capability is trending up about 13.7 Epoch Capabilities Index points per year at an R² of 0.95, which puts small models at GLM-5.2's replication level by the end of 2026. Scaled against the Kimwolf botnet's two million infected devices, the team estimates roughly 133,000 replicas, about 80,000 tokens per second, and at most 40 H100-equivalents of compute. Replicating a 9B copy takes one to two days, so the doubling time is days, not minutes. The authors call this "signs of life" rather than a rogue replicator, and The Register argues Anthropic is again publicizing a threat that does not exist yet.

The defensive argument is the useful part, and it is unglamorous. The thing that makes an AI worm different from ordinary malware is that it can work out unfamiliar hardware on its own, and the thing that makes it hard to kill is that there is no central server to switch off. Anthropic's suggested answers are ISP-level quarantine with user notification, which cleared 92% of Mirai infections within 14 days versus 74% without notification, manufacturer security baselines to shrink the pool of easy targets, and network-signature detection. Boring, and available now.

Read more.

Gemini 4 Argon is out of the cyber-defender sandbox, and cheap until December

Google made Gemini 4 Argon broadly available on Tuesday. When it was announced on September 30 it was restricted to approved cybersecurity defenders through the Fairwind Program with no availability date. It is now live on the paid Gemini API, in AI Studio, in Antigravity, and on eligible Gemini Enterprise plans, with AI Pro and Ultra subscribers getting it in the Gemini app and AI Mode in Search. Android developers reach it through Android Studio and Firebase AI Logic. Antigravity for individuals still requires the Ultra tier.

Argon runs $2 per million input tokens and $10 per million output through November 30, with cached input discounted 95% to $0.10, then jumps to $4 and $20 on December 1. The intro rate puts it level with Claude Sonnet 5.5 and GPT-6.1 Sol; the December rate puts it level with Claude Opus 5.5 and Meta's Muse Wonder. Google also describes a unified ceiling of up to 1M input and output tokens through a continuation protocol, which is a different claim from the 1M output limit in the September announcement, and it has not clarified how that relates to the length of any single response.

Google's own numbers have Argon at 77.9% on DeepSWE v1.1 against 74.2% for Opus 5.5, 19.6% on Harvey Legal Agent against 5.4% for GPT-6 Astra, 51.3% on Zapier AutomationBench, 87% on Embodied Reasoning QA, and a 0.7% attack success rate on Gray Swan's indirect prompt injection suite over 15 adaptive attempts. It loses Terminal-Bench 4.0 to Opus 5.5, 57.4% to 66.4%. Independent reads split: Vals has Argon first on its index at 68.90%, ahead of Sonnet 5.5 and Opus 5.5, while Artificial Analysis puts it at 53, tied with GPT-6 Astra and Claude Fable 5.1 and five points behind Opus 5.5 at 58.

Google says more than 100,000 employees use Argon monthly through an internal Antigravity build, with agentic code submissions up 35% and post-submission rollbacks down 30% over four weeks. All of that is self-reported. Reuters frames the launch as Google trying to recover ground after months of delays and a DeepMind leadership shakeup, which reads about right: Argon is competitive, priced to be chosen, and not the clear frontier leader on independent indices.

Read more.

Scion is Anthropic's replacement for git, and the numbers behind it explain why it exists. Daily merges into its monorepo went from about 2,500 in May to more than 30,000 by September, which is what happens when agents start opening most of the pull requests. Seven engineers built it in five months, and Anthropic claims a 200x reduction in version control infrastructure footprint. Simon Willison's notes are the fastest way to get the design argument. Not something you can adopt, but a good look at which tools break first under agent-scale throughput.

Intel's GPU D270 launched formally at the OCP Global Summit, and the bet is interesting: up to 480GB of LPDDR5X instead of HBM, 1.3 PFLOPS of FP8, air-cooled at 350W, and no 3D graphics pipeline at all. The pitch is price per token on large-model inference rather than peak throughput, aimed at serving models up to roughly 300B parameters without an HBM supply fight. Customer sampling is underway with broader availability targeted for 2027, so this is a 2027 purchasing decision, not a 2026 one.

NVIDIA contributed its Vera Rubin rack and compute tray designs to the Open Compute Project through MGX, alongside an Open Data Center Specification v0.7 co-authored with AMD, Google and Meta. MGX 1.2 adds blind-mate liquid cooling, a liquid-cooled 5,000A busbar and support for racks up to 250kW. NVIDIA also open-sourced four DSX OS pieces, including a Cluster Readiness Engine that stress-tests GPU clusters before you commit a training run to them. No Vera Rubin shipping date.

Google DeepMind and Google Cloud are opening AlphaFold 3 and other science models to up to 200 European researchers with $3 million in compute credits, backing the EU's RAISE program. Worth a look if you are at a European institution, and worth noting as the shape these access deals are starting to take.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.