Latest / UpNext AI / Cyera’s $1B AI Security Bet, a $410M Compute Deal, and the Agent Benchmark Problem | UpNext AI – July 29, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, July 29th, 2026, and here's what matters in AI today. We start with enterprise security, where Cyera has agreed to acquire Oasis Security for approximately 1 billion dollars. According to TechCrunch, the deal is expected to be paid mostly in cash, with the remainder in Cyera shares. Cyera recently raised 600 million dollars at a 12 billion dollar valuation, and this would be its third acquisition of the year. Why this one matters is the target. Oasis focuses on non-human identities, primarily AI agents. And that gets at a real enterprise problem: as companies deploy more agents, they need systems that can monitor those agents’ behavior and control what software and data those agents can access. TechCrunch says Cyera plans to fold Oasis into a unified identity and data security platform. So this is not just another cybersecurity deal. It’s a sign that securing AI agents is quickly becoming its own buying category, with real M and A money behind it. The broader takeaway is simple: once agents start acting inside real business systems, identity and access control become much more important. This acquisition is a pretty clear signal that vendors think that market is arriving fast. From security infrastructure to raw infrastructure, Recursive Superintelligence has signed a 410 million dollar compute deal with Amazon Web Services. TechCrunch reports that Recursive emerged from stealth in May with 650 million dollars in funding, and the new AWS agreement is multiyear. The company is focused on self-improving AI systems, and the story says that focus pushes much more of its budget into compute instead of headcount and operations. That framing is the key part here. Recursive is trying to automate more of its own product development process, and its founder Richard Socher told TechCrunch that for the company, it is less about headcount and more about agent count. There’s also no investment component from Amazon here. This is a compute agreement, not one of those hybrid strategic tie-ups we’ve seen elsewhere. But AWS says part of the arrangement is co-developing infrastructure purpose-built for companies with these kinds of needs. So the second big theme today is that frontier ambitions increasingly look like infrastructure purchasing decisions. If a startup thinks self-improving systems are the path to useful products, it may spend hundreds of millions on compute before it spends the equivalent on people. Now to the research section, and this one is about measuring agents rather than building them. A new paper from earlier this week is called Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation. The paper argues that agent evaluation is still badly fragmented across tasks, scaffolds, verifiers, and scoring rules, which makes benchmark results hard to compare across the field. The authors say Messier spans 30 benchmarks and standardizes records around the model, scaffold, environment, task, verifier, and aggregation rule. In plain English, a scaffold is the extra system wrapped around a model to help it act like an agent, and the paper’s point is that those wrappers can change results a lot. The dataset is large, and the paper says it includes 957,253 records covering 714 agents, 11,891 tasks, and 74,205 verifiers. Using that corpus, the authors say frontier progress looks uneven depending on the benchmark type. They report that function calling looks relatively saturated, programming is improving fastest, and enterprise workflows remain the hardest category. One especially useful point: the paper says scoring choices can distort rankings. In multi-verifier tasks, strict all-pass scoring can hide progress and even reshuffle which agents look best. Bottom line: if every agent benchmark uses different rules, it is hard to know which system is actually better. Messier is an attempt to give the field a more common yardstick. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... First, Nvidia is pushing its Jetson line as a compact platform for edge AI and robotics. In a company blog post, Nvidia highlighted Jetson modules and developer kits for robots, autonomous machines, and on-device AI projects, with examples ranging from student builds to research systems. One named spec that stood out: Jetson Orin Nano Super is described as delivering 67 trillion operations per second of AI performance. Next, PayPal says it could still consider a takeover offer if it creates more value for shareholders. That came after better-than-expected Q2 results, according to TechCrunch. The story says Stripe and Advent had previously offered about 53.4 billion dollars, or 60 dollars and 50 cents per share. TechCrunch also cites analysis valuing PayPal closer to 70 dollars per share, while the stock was trading around 58. Also today, Granola is launching an Apple Watch app. TechCrunch says the AI note-taking company wants users to be able to record meetings and take notes without pulling out a smartphone, especially for in-person meetings like walking one-on-ones. Granola raised 125 million dollars earlier this year. And finally, the Financial Times reports that AI risk is dragging on a 5 billion dollar refinancing tied to Thoma Bravo-backed software, with cybersecurity firm Proofpoint facing higher borrowing costs and tighter covenants. We only have limited sourced detail here, but the headline itself is notable: lenders appear to be pricing AI-related risk into large software financing more aggressively. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!