In partnership with
In Today’s Issue:
🕵️ Someone bugged OpenAI, and AI 2027 saw it coming
🎮 Epic rebuilds its Disney shooter with AI
🗂️ Claude writes so much code that Anthropic outgrew git
🔁 Jack Clark moves his self-improving AI forecast to 2027
✍️ A detector that catches outsourced thinking
✨ And more AI goodness…
⚡ The Signal
The most valuable thing an AI lab owns now fits into a conversation, and someone tried to record it.
OpenAI found a hidden recorder behind the TV in an executive conference room, used its own models to trace it, and the FBI arrested an engineer. Read next to the AI 2027 scenario, the case feels less like a one-off: that forecast expected lab secrets to leak through people in ordinary offices, and it put this chapter in 2026 and 2027. The rest of today's issue shows why the stakes keep rising. Anthropic merged more than 30,000 code changes in a single September day, Epic is betting a Disney game on AI-first development, and Anthropic co-founder Jack Clark now expects self-improving AI to become possible in 2027, right on the AI 2027 timeline. When the work itself speeds up this fast, the ideas in the room become the thing worth stealing.
All the best,
Kim Isenberg
(Epic Games via Kotaku)
🎮 Epic Delays Its Disney Shooter to Rebuild It With AI
Epic Games has delayed its first Disney game, a shooter with Marvel and Star Wars characters planned for March 2027, and told the team to finish it with AI doing most of the work. According to Bloomberg, developers now follow an "AI first" model in which tools such as Anthropic's Claude write the code while humans steer; one director wrote "No more hand-written code." Epic says the push is led by its own developers, while some staff reportedly reacted with shock and doubt. Disney invested $1.5 billion in Epic in 2024.
👉 tl;dr: A major studio is betting a Marvel and Star Wars game on AI writing the code, a high-profile test of whether AI-first production can rescue a project that fell behind.
(Anthropic)
🗂️ Claude Writes So Much Code That Anthropic Outgrew Git
Anthropic averaged about 2,500 code changes a day in May and merged more than 30,000 on a single day in September, so much AI-written code that git, the standard tool developers use to track changes, became the bottleneck. A team of seven built Scion, an in-house version control system designed for AI agents that are always online, and moved the company's giant shared codebase onto it in five months. Code checkouts now take seconds instead of minutes on infrastructure 200x smaller than the old git setup. The engineers say Claude now writes code better than they do; their job is often to tell it "no."
👉 tl;dr: AI coding has grown so fast inside Anthropic that it had to rebuild a basic developer tool, and it expects other companies to hit the same wall within six to 12 months.
(Getty Images via NPR)
🔁 Anthropic Co-Founder Jack Clark Now Expects Self-Improving AI in 2027
Jack Clark, co-founder of Anthropic, now expects AI that can fully improve itself, known as recursive self-improvement or RSI, to become possible a year sooner than he previously thought. "I now think it's more likely we'll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028," he writes in his Import AI newsletter. His reasons: AI systems are making creative-seeming math breakthroughs, and on the new TasteVal benchmark their research taste has doubled every three months since December 2025 (more in today's Graph of the Day). That lands close to the AI 2027 scenario from today's lead story, which puts self-improving AI in mid-2027.
👉 tl;dr: An Anthropic co-founder now sees self-improving AI as most likely possible in 2027, a year earlier than he thought and in line with the AI 2027 scenario.
The agentic era needs a different CRM. That’s Attio.
Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.
🎬 Watch This
Today I'm on the other side of the microphone. In a German-language conversation with strategist and creative director Ole Tillmann, premiering today at 14:00 CEST, we take on the question behind every AI debate: will AI replace us? I explain why serious forecasts range from Nobel laureate Daron Acemoglu's roughly 1% boost to economic growth to the 20 to 30% a year that Elon Musk and Dario Amodei have floated, why I think AI is the first technology that can replace people rather than only speed them up, and what still makes me optimistic.
"AI safety is in our control. … Safety is an engineering problem."
– Jensen Huang, founder and CEO of Nvidia, at a press conference on October 13
Huang's case is that AI risk is something the people who build it can test and fix. Today's lead story adds the other half of that control problem: in OpenAI's case, the threat was a person with a hidden recorder, and the company's own models helped track him down.
Moonshot AI's road to a Hong Kong IPO is getting bumpy. Bank of America is no longer among the banks that could work on the Kimi maker's listing as the company faces US scrutiny, Bloomberg reports, while Reuters says its founders are trying to settle a shareholder dispute over a stake that could be worth more than $700 million.
Someone Bugged OpenAI, and AI 2027 Saw It Coming
The Takeaway
👉 OpenAI found a hidden recorder in an executive conference room on September 30; the FBI arrested an OpenAI engineer on October 9.
👉 OpenAI says its own models helped analyze the device and identify who planted it.
👉 OpenAI sees no sign that model weights, source code or customer data were taken; the charge is an allegation, not a conviction.
👉 The AI 2027 scenario expected this pressure on schedule: secrets simple enough to say out loud, guarded by tech-company rather than military-grade security.
Someone tried to listen in on the room where OpenAI's leaders make their biggest decisions. On September 30, OpenAI staff found a recording device wrapped in toilet paper and plugged into the back of a TV in an executive conference room in San Francisco, according to a federal complaint. OpenAI says it used its own models to work out what the device was and where it came from, which led to an engineer who had joined in July 2025. The FBI arrested Abdulrahman Alfozan, 35, on October 9. Prosecutors charge him with attempting to intercept oral communications; he faces up to five years in prison if convicted.
The complaint reads like a spy novel. Security cameras showed him near a room he had no access to, a chatbot account linked to his spouse asked how to build "a continuous audio recorder" and encrypt "the SD card," and a second recorder turned up in his bag. FBI Director Kash Patel says the alleged aim was to "secretly capture sensitive communications about the company's super intelligence technology." Prosecutors have not said who, if anyone, was meant to receive the recordings, and OpenAI says it sees no sign that model weights, source code or customer data were taken.
(OpenAI via TechCrunch)
This is where AI 2027 looks uncomfortably prescient. The scenario, published in April 2025 by Daniel Kokotajlo's AI Futures Project, placed the lab-espionage chapter in mid-2026 and warned that algorithmic secrets "simple enough to relay verbally" would stay exposed because lab offices had "security more typical of a tech company than a military operation." No government has been linked to this case. Yet the official reaction already sounds like the scenario: the FBI calls protecting American superintelligence innovation one of its "highest counterintelligence priorities," and the Director of National Intelligence framed the case as part of the "global super intelligence race."
(AI Futures Project, AI 2027)
Why it matters: The most valuable thing a frontier lab owns now includes what its leaders say behind closed doors, and Washington has started treating its protection as national security. Labs should expect security that looks more like a defense contractor's than a startup's, and OpenAI just showed that AI itself can be part of that defense.
Sources:
(P-Zero Research, TasteVal)
The chart: P-Zero Research's new TasteVal benchmark measures research taste: how well a model designs experiments and reads their results on eight real AI-development tasks, scored by how much compute it needs compared with human experts (the dashed 1x line). Progress was slow from GPT-4 in 2023 through 2025, then the slope steepened: since December 2025, scores have doubled every 3.0 months. Claude Opus 5.5 now sits at 2.30x the expert baseline, ahead of Fable 5.1 and GPT-6 Astra, at roughly 1/30 of the experts' cost per attempt.
The lesson: Research taste is the skill that decides how quickly AI could start improving AI, and this curve is one reason Jack Clark (Story 3) now expects that in 2027. In a naive extrapolation the authors stress is not a forecast, this trend pulls the AI Futures Project's median date for superintelligence from mid-2030 to late 2028. That is the team behind AI 2027, and the reason today's lead story is about more than one bugged room.
The caveat: The tasks are quick and cheap to check, the kind labs find easiest to optimize for, and the human baseline does not include the world's top researchers, so the curve may overstate real progress.
✍️ The AI Detector That Catches Outsourced Thinking
⚡ Bottom line
Today's AI detectors only check who wrote the words; the new IdeaLens checks whether the ideas came from AI.
💡 Why it matters
Letting ChatGPT plan an essay and typing it yourself fools today's detectors; IdeaLens catches most of these cases.
🔎 What it means
Using AI to polish your own thinking gets easier to defend; outsourcing the thinking itself gets harder to hide.
A student can fool today's AI detectors with a simple trick: let ChatGPT come up with the essay, then type it out in their own words. Researchers tested exactly that. People wrote 50 short stories by hand, each based on a plan from OpenAI's GPT-5.6 Sol, and Pangram 4, one of the best-known detectors, flagged only 8% of them as AI. It also works the other way around: bring your own ideas, let AI polish the wording, and the same detector flags you.
The reason is simple. Detectors like Pangram check who wrote the words. So do the invisible watermarks Anthropic announced for Claude's text in August, following Google's SynthID: they sit in the words a model generates, so an essay typed by hand from a Claude plan carries no mark at all. A new tool called IdeaLens, from the University of Maryland, Google DeepMind and Simon Fraser University, checks whose ideas they are. It reduces a text to its skeleton, the claims, the events and the structure, and judges only that.
(Superintelligence chart; data: Rajendhran et al., arXiv)
The difference is large. On the hand-written stories based on AI plans, IdeaLens flagged 68% as AI, against Pangram 4's 8%. On texts that AI wrote from detailed human plans, it flagged only 7%, where Pangram 4 flagged 92%. IdeaLens goes after outsourced thinking and largely leaves AI-polished writing alone.
(TheWrap)
Editors are already fighting over exactly this. In August, Pangram flagged a Wall Street Journal op-ed by billionaire investor Stanley Druckenmiller as 100% AI-written, and he replied: "Of course I used AI." The Journal's opinion editor, Paul Gigot, defended him, saying what matters is whether a piece "reflects an author's original argument." IdeaLens is built for that question: it rated Druckenmiller's ideas 98.3% likely to come from AI, an estimate the authors stress is not proof. The researchers have released their models and data, so schools and publishers can try it themselves.
Most AI Tools Weren't Built for This
B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.