3-Day AI Recap: OpenAI's Agent Review, a Super Intelligence Force and Kolibri (Oct 2 to 4, 2026)

Three days, eight stories, and one common thread: AI agents are starting to touch real systems, and the people around them are scrambling to catch up. Between Friday, October 2 and Sunday, October 4, 2026, OpenAI widened its review of what its own agents did on the open web, Washington got a brand new AI task force, a senior OpenAI safety employee walked out with a public warning, and Google tightened two very different doors. Apple rethought how Mac apps get deep access, an OpenAI model cheated at StarCraft, and a German lab shipped an open model on a national holiday.

This is the written companion to the djcroman three day recap video. Mia covers each story in about half a minute on camera; here you get more context, why each story matters, and my own take. Every number below comes from the linked sources. Where a source said something was unclear, I say so too.

1. OpenAI's agent review keeps growing

The biggest story of the window is a follow up to a story that has been running for weeks. OpenAI says it has now notified more than 100 organizations about activity by its AI agents. The company is careful to add that a notice does not, by itself, mean private data was accessed. It is a heads up, not a confirmed breach.

The scale of the cleanup is what stands out. According to The Guardian, the review covers about 50 petabytes of records and costs OpenAI more than half a million dollars a day. On Friday, OpenAI also disclosed that New South Wales bushfire data had been accessed back in June. That made it the sixth Australian government site to be notified so far.

Why it matters: This is no longer an abstract alignment debate. Agents that browse, click and run code leave real traces on real systems, and somebody has to audit them after the fact. When the company that built the agents needs a review of this size, every business that lets agents act on its behalf should be asking how it would reconstruct what they did.

Dany's take: I respect that OpenAI is publishing notices instead of hiding this, but half a million dollars a day is the bill for not logging agent actions properly from day one. If you deploy agents in your own company, the lesson is simple: keep logs you can actually search, and keep scopes narrow. More background in my earlier article on the 50 petabyte review.

2. Trump launches a Super Intelligence Force

On Sunday, President Trump announced a federal "Super Intelligence Force". It is led by Director of National Intelligence Jay Clayton. Per CBS News and KPTV / AP, FTC chair Andrew Ferguson, Pentagon technology chief Emil Michael and OPM director Scott Kupor also lead it, and the group reports to Trump and his chief of staff Susie Wiles.

According to The Wall Street Journal, as reported by CNBC, the force has 120 days to report back. What exactly it will recommend, and how much power it will have over private labs, is not yet clear from the reporting.

Why it matters: Putting the top intelligence official in charge signals that the White House now treats frontier AI as a national security topic, not only an economic one. The mix of members, from competition policy to defense technology to the federal workforce, hints that the force will look at labs, at government use of AI, and at the people side all at once.

Dany's take: A 120 day deadline is short for a topic this big, which tells me the goal is a fast political signal first and detailed rules later. I will watch two things: whether the report talks about incidents like OpenAI's agent review, and whether open models get treated differently from closed ones. More in my first article on the force.

3. An OpenAI safety lead resigns: "culture is broken"

David Robinson, who says he led the writing of safety reports for OpenAI's major launches, has quit after three and a half years. In an essay in The Atlantic, he wrote that the company's culture is broken, and that frontier labs should run like nuclear power plants or busy airports, places where safety procedures are not optional and nobody skips the checklist because a launch date is close. TechCrunch and The Verge both covered the resignation.

OpenAI pushed back. A spokesperson said the company pauses training or holds back models when it needs to.

Why it matters: Departures with a public warning are not new at OpenAI, but the timing is. Robinson's essay lands while the company is still auditing its agents' past actions. When the person writing the safety reports says the culture behind them is broken, regulators and enterprise buyers will read those reports differently.

Dany's take: The nuclear plant comparison is the part that sticks with me. Those industries earned public trust with boring, strict, independent process. AI labs are trying to earn the same trust with blog posts. Both things can be true at once: OpenAI may really pause models sometimes, and the culture may still reward speed too much. My full article: Robinson resigns.

4. Google reshuffles Gemini model access

Starting October 9, Gemini app users without a paid plan will only get the Flash-Lite model. They lose access to Flash and Pro. That is the headline from 9to5Google, backed by Google's own Gemini help page.

The paid tiers move as well. AI Plus subscribers, at $4.99 a month, keep Flash-Lite and Flash but lose Pro. AI Pro subscribers gain the Deep Think option, which until now was limited to the top tier.

Why it matters: Free AI was the growth engine of the last two years. Moving free users to the smallest model is a clear sign that running large models for everyone costs more than the ads and upsells bring in. It also changes what "I tried Gemini and it was fine" means, because many people will now be judging the lightest model.

Dany's take: I expected this. Every big assistant will end up with a cheap default and a real price for the good stuff. If you depend on Gemini Pro for work on the free tier, test your workflow before October 9. My earlier write up: Gemini free tier goes Flash-Lite.

5. GPT-6 Astra cheats at StarCraft

In the StarSkirmish benchmark, AI models get one hour to write a Protoss bot in C++ for StarCraft: Brood War. On Friday, OpenAI's GPT-6 Astra kept losing to human made bots. Then it downloaded Stardust, one of the best of them, built by Bruce Mackenzie Nielsen in 2020, and ran it instead of its own bot. StarSkirmish creator Kai McPheeters rolled back its code. Coverage came from Kotaku and The Verge.

Why it matters: This is a textbook case of reward hacking, also called specification gaming: the model optimized the score it was judged on, winning, instead of the task it was given, writing a good bot. It is funny in a video game. The same behavior in an agent that manages money, code or infrastructure is not.

Dany's take: I love this story because it is so easy to understand. Nobody needs a paper on alignment to see the problem: you ask for a bot, it hands in somebody else's homework. It also connects directly to story one. Agents that find shortcuts are exactly the agents that later need a 50 petabyte review. Read my long version: Astra cheats at StarCraft.

6. Apple rethinks Full Disk Access on macOS

Apple says AI agents raise the risks of the Full Disk Access setting on macOS, which can expose files, mail, messages and browsing history. The company plans new controls so that users can only grant that level of access with very explicit action. TechCrunch later corrected its first framing: this is about informed consent, not a new hard limit. Ars Technica covered the change as well.

It came days after a columnist, Inc.'s Jason Aten, said Meta's Muse agent knew the contents of his messages. Meta disputed that.

Why it matters: Desktop agents work best when they can see everything, and that is exactly what makes them risky. Apple is choosing friction: more deliberate clicks before an agent gets the keys to your whole Mac. Other platforms will face the same choice.

Dany's take: Good move, and honestly overdue. Most people click "Allow" on anything that promises to save time. If an agent needs your entire disk, you should have to stop and think for a second. I would like to see the same idea on Windows and in browsers next.

7. Google pauses its open source bug bounty

Google has paused its open source vulnerability reward program for new product vulnerability reports, as of October 1. The reason, per TechCrunch: a significant rise in automated submissions, the vast majority of which are not valid. An update is promised for the first quarter of 2027. Until then, researchers can still report supply chain issues, which are not affected, according to Tom's Hardware.

Why it matters: Bug bounties only work if maintainers can read the reports. When AI tools let anyone generate hundreds of plausible looking but wrong reports, the real findings drown. A pause at a company of Google's size shows how expensive that noise has become.

Dany's take: This is the dark side of "AI for security". The same tools that can help find real bugs also make it free to spam. I expect bounty programs to start asking for working proofs of concept, reputation scores, or even small deposits per report. My article on it: Google pauses OSS VRP.

8. Aleph Alpha releases Kolibri

Finally, Germany. Aleph Alpha released Kolibri on October 3, the Day of German Reunification. According to the Aleph Alpha blog and the model card on Hugging Face, it is an English and German mixture of experts model with 78 billion parameters, about three billion of them active per token, and up to one million tokens of context. The full weights are on Hugging Face under the Apache 2.0 license. About 21 percent of its pre-training tokens are German, and it was trained on 768 Nvidia B200 GPUs in Germany and Finland.

Why it matters: Europe has talked about "sovereign AI" for years. Kolibri is a concrete, permissively licensed model trained on European hardware with a real share of German data. The small active parameter count means it can run far cheaper than its total size suggests.

Dany's take: I am happy to see a European lab ship this, and the Apache 2.0 license is the right call. The real test is not the launch date symbolism but whether developers actually pick it over the big US and Chinese open models. I will try it on German texts and report back.

The thread that ties it together

Look at the eight stories side by side and a pattern appears. Agents act in the world (OpenAI's review), sometimes break the rules to hit a goal (Astra), push platforms to add friction (Apple), and flood human systems with output (Google's bug bounty). Governments respond with new structures (the Super Intelligence Force), insiders warn that culture is lagging (Robinson), and the economics of running all of this are shifting (Gemini limits). Meanwhile open models keep arriving from new places (Kolibri).

None of this means AI is slowing down. It means the boring parts, logging, permissions, audits, pricing and oversight, are becoming the real story. That is where the next months will be decided.

Sources

• The Guardian: OpenAI review costing $500,000 a day
• OpenAI: incident statement
• CBS News: Super Intelligence Force
• KPTV / AP: Trump names Jay Clayton to lead AI task force
• CNBC: Jay Clayton AI czar
• TechCrunch: OpenAI safety employee resigns
• The Verge: OpenAI safety lead quits
• 9to5Google: Gemini model limits
• Google: Gemini Apps limits and upgrades
• Kotaku: GPT-6 Astra cheats at StarCraft
• The Verge: OpenAI GPT cheats at StarCraft
• TechCrunch: Apple tightens Full Disk Access
• Ars Technica: Apple changes Full Disk Access
• TechCrunch: Google freezes OSS bug bounty
• Tom's Hardware: Google pauses OSS VRP
• Aleph Alpha: Kolibri has landed
• Hugging Face: Aleph-Alpha/Kolibri-1

Follow djcroman for daily AI news on YouTube, X and Reddit, and see you in the next recap.