What would GPT-6 Cyber have to prove before you'd pay more for its exploit-hunting? This is AI Daily Briefing. Also today: Dropbox's infrastructure shopping list, and how much choice Google is actually giving schools. First, OpenAI. What evidence would make Cyber worth trusting on defense?
Carl Franzen, writing in VentureBeat:
OpenAI released GPT-6 Cyber today, a specialized security model that the company says can uncover and exploit more software vulnerabilities while more reliably respecting the limits researchers impose. For enterprise security teams, the launch offers another tool for testing defenses and developing fixes, but access requires approval and its published token prices exceed those of competing cyber models from rivals Anthropic and Google.
GPT-6 Cyber costs twenty dollars per million input tokens, a hundred per million output tokens. At that output price, I'd want a very short victory speech after it finds a bug. VentureBeat puts those rates above Anthropic and Google's competing cyber models. And having the budget doesn't get you access. OpenAI's starting with selected customers in Daybreak Red; applicants submit their intended defensive workflows for review. OpenAI gets to sell the capability and decide whose security work qualifies. I'd still trial it. Cached input drops to two dollars per million tokens, which could help when you're repeatedly examining the same code. But a hundred-dollar output rate means I'd put a spending cap on that trial before giving it a repository.
Crunchbase News, with Mary Ann Azevedo:
Dropbox announced Thursday plans to acquire Tigris Data for $350 million in a transaction the company told Crunchbase News would mark its largest acquisition since its founding in 2007. Founded by former Uber engineers, Sunnyvale, California-based Tigris is building a network of localized data storage centers that it claims can meet the distributed compute needs of modern AI workloads.
Dropbox's new DBX Agent Cloud starts with standalone object storage. Sensible plumbing underneath an extremely fashionable name. The planned $350 million Tigris acquisition gives it a route to developers running applications across multiple clouds. And Dropbox told Crunchbase News this would be its biggest acquisition ever. That's a substantial bet on selling infrastructure to developers—a different sales job from getting businesses to pay for shared folders. Tigris says its localized storage network can serve distributed AI workloads. I'm interested: moving data around can punish both response times and the bill. I'd test that claim with an application spread across clouds before moving anything important. And the eightfold annual growth figure? That came from Tigris CEO Ovais Tariq at its October 2025 funding round. Keep the date attached. Dropbox's $350 million bet deserves a fresher yardstick.
From Ryan Whitwam at Ars Technica:
The company clarified in a tweet that it will allow schools to disable Gemini apps while leaving search active for students. If they don’t want any AI, they can leave it all off. Essentially, Google is giving schools that want to completely block AI the ability to go all the way, even removing search from student accounts. A Google spokesperson tells Ars that the company has not made any changes to the new policy it announced this week.
Schools can leave Google Search on and switch Gemini apps off. Ars got that clarification after the backlash, and it matters: a district rejecting the chatbot doesn't automatically lose search for logged-in students. But schools that want search without AI still don't get that choice. Google put Search inside its AI portfolio, then described the new policy as giving schools more control. Administrators get separate switches for Gemini and Search; Google still decides what comes inside Search. I'll give Google credit for those separate switches. They're useful. But its spokesperson tells Ars the policy hasn't changed, which means this whole fight erupted over explaining the controls schools are supposed to use. Pretty consequential place to be vague.
When a company says its AI is better at finding software exploits, why shouldn’t we treat a high benchmark score as proof it will make defenders safer? What would actually count as convincing evidence? The first test is whether the model can work on vulnerabilities it has not effectively seen before. Researchers studying AI and nonpublic vulnerabilities warn that public benchmarks may contain already published advisories, exploits, and fixes, so a strong score can reflect prior exposure rather than an ability to uncover new flaws. A more meaningful evaluation would use held-out codebases, then measure not just whether the system flags a bug, but whether people can verify the finding, reproduce it, and safely fix it. Gonzalo Gonzalez-Pumariega and colleagues make that distinction in CWE-bench: they say a working exploit does not itself show that an agent can make software safe, because an attacker needs only one successful vulnerability while a defender has to find and fix them all. Their private benchmark includes 120 held-out audit-and-repair tasks across 73 weakness types, eight programming languages, and the OWASP Top 10 2025 categories. And an independent reproducibility study of 104 papers points to another practical bar: AI-generated security artifacts may be runnable or produce a signal without being semantically confirmed as valid. So even if the model really does find more bugs, could that create a defense problem by flooding the people who have to decide what is real and urgent? Yes. TechTarget reporter Ben Lutkevich reports that noisy AI-generated vulnerability discoveries can bottleneck human triage, and that only a small share of initial Mythos findings entered Anthropic’s disclosure ledger. The risk is that defenders spend scarce time sorting false alarms while the same capability can widen the attack surface; the useful metric is validated, remediated vulnerabilities, not raw findings. What to watch for is independent testing that reports false positives, verification rates, repair success, and performance on genuinely unseen software.
Have feedback, a story idea, or a correction? Email us at aidailybriefing at lantern podcasts dot com. Your notes help make AI Daily Briefing more useful every day.
Looking ahead, OpenAI says GPT-6 Cyber access will expand to all approved Daybreak Red customers over the coming weeks, while Dropbox expects its $350 million Tigris Data acquisition to close in the fourth quarter of 2026. Links to every story are in the show notes, so take a look at the ones that caught your attention. Thanks for listening, and we’ll see you next episode. This is a Lantern Podcast.