This Week in AI — June 26

For two years the pitch was that AI could find the problem. This week the pitch changed to AI fixing it, paying for it, and quietly poaching the people who built the field in the first place. Five things worth your attention from the past seven days.

OpenAI's cyber model stops hunting bugs and starts patching them

On June 22 OpenAI shipped the full version of GPT-5.5-Cyber and folded it into a security program called Daybreak, plus a "Patch the Planet" effort aimed at open-source projects. The model posted 85.6% on the CyberGym benchmark, up from 81.8% for plain GPT-5.5. The interesting part isn't the score. It's the verb. OpenAI says its Codex Security tool has already scanned over 30 million commits across 30,000-plus codebases, with more than 500,000 findings auto-determined as fixed.

Why it matters: the whole AI-security story has been about detection. Finding the hole was the easy half. Closing it without breaking production was the hard half, and that's where the human hours lived.

Operator take: when the machine can find the problem and propose the fix, the scarce skill shifts to verification. Same lesson as every support automation we run. The bot drafts, a person signs off.

OpenAI: Scaling Trusted Access for Cyber

SpaceX rents out its supercomputer for $150M a month

SpaceX signed a compute deal with open-source startup Reflection AI worth up to $6.3 billion. Reflection pays $150 million a month, starting July 1, for Nvidia GB300 chips at the Colossus 2 data center near Memphis. CNBC and Bloomberg both confirmed it on June 22. Colossus has now signed Anthropic, Google, and Cursor, turning a rocket company's side project into a commercial compute landlord.

Why it matters: the bottleneck in AI right now isn't ideas, it's chips and the power to run them. Whoever owns the racks sets the meter.

Operator take: you don't need a Memphis data center, but the principle scales down. The cost of intelligence is becoming a line item you negotiate, not a fixed price you accept. Watch your per-call AI spend the way Reflection watches its monthly invoice.

CNBC: SpaceX signs computing power deal with Reflection

Google's brain trust keeps walking out the door

John Jumper, the Nobel laureate behind AlphaFold, announced on June 19 he's leaving Google DeepMind for Anthropic after nearly nine years. He follows Noam Shazeer, a co-author of the original Transformer paper, who left for OpenAI days earlier. By June 23 Fortune was openly asking whether DeepMind can stay at the front of the race. Google shares dropped more than 5% on the news.

Why it matters: frontier AI is still a small-numbers game. A handful of people move the line, and right now they're moving toward Anthropic and OpenAI.

Operator take: talent concentration is a tell. When the best researchers vote with their feet, they're forecasting where the next breakthrough ships. Worth tracking who's hiring, not just who's announcing.

CNBC: John Jumper to leave Google DeepMind for Anthropic

Anthropic puts a coworker named Claude inside Slack

On June 23 Anthropic launched Claude Tag, an agent that lives in a Slack channel. Tag @Claude with a request and it breaks the job into steps, runs them with the tools it's given, and replies in the thread. One Claude per channel, shared by everyone, so anyone can pick up where the last person left off. It's in beta for Team and Enterprise plans and replaces the old Claude in Slack app, with a switchover set for August 3.

Why it matters: this is the agent-in-the-workflow bet made concrete. Not a separate tab you visit, a teammate you @-mention where the work already happens.

Operator take: the winning automation isn't the one with the best demo, it's the one that meets people in the tool they already live in. For most teams that tool is Slack, email, or the helpdesk. Build there.

Fortune: Anthropic launches Claude Tag for Slack

IBM and OpenAI team up to defend at machine speed

Also on June 22, IBM and OpenAI announced a partnership to bring frontier models into enterprise cyber defense, the idea being to match the speed of AI-built attacks with AI-built defenses. It lands the same week OpenAI expanded Daybreak, which is not a coincidence. The pitch to a CISO is blunt: the attackers already have these tools, so your defense can't run at human pace anymore.

Why it matters: enterprise security buyers move slowly, and a name like IBM is how frontier AI gets through procurement. Distribution beats novelty.

Operator take: the same logic applies far below the enterprise tier. The competitor answering leads in five minutes with AI isn't waiting for you to catch up. Speed is the moat, and it compounds daily.

IBM Newsroom: IBM and OpenAI bring frontier AI to cyber defense