What AI's Most Innovative Thinkers are Saying

Welcome to AI Confidential, your biweekly breakdown of the most interesting developments in confidential AI.

Today we’re exploring:

• Our best 2025 podcast highlights

• Why IT decision-makers are losing sleep over AI

• Open source projects you need to know about

Also mentioned in this issue: Reuven Cohen, João Moura, Marco Palladino, Sriram Raghavan, Vijoy Pandey, Mark Russinovich, Atin Sanyal, CrewAI, Kong, IBM, Cisco, Microsoft, Galileo, Temporal.io, Infosys, DeepSeek, IT Pro, StepFunAI, and Skywork.

Let’s dive in!

If you haven’t checked out our AI Confidential podcast, you’re missing out. 

Over the past 10 episodes, we’ve heard from some of AI’s most innovative builders and thinkers who are reimagining exactly what security means in the AI era. 

Here are a few of our favorite takeaways:

1️⃣ AI agents are moving into the enterprise

If you’re reading this newsletter, you know about the power of agentic AI.

Turns out, the experts feel the same way—because we’ve been talking about agents for months on the pod:

✔️ Agentic Foundation’s Reuven Cohen (AKA rUv) explained how he uses agents to generate 10 million lines of code in a year, revolutionizing his workflows and productivity.

✔️ CrewAI’s João Moura detailed how he’s making it easy for developers to build human-like agents through open source, including massive global enterprises that use their agents to streamline work.

✔️ Kong’s Marco Palladino warned that agents are fast becoming the new API client—delivering productivity and a new software challenge, but also coming with new risks.

2️⃣ Enterprise AI is being redefined

As AI options become more and more complex, it’s been eye-opening to see how the most prominent enterprise vendors handle the shift:

✔️ IBM’s Sriram Raghavan argued that smaller, specialized models are driving the next wave of enterprise innovation, which changed how we viewed the unquestioned power of LLMs.

✔️ On the other hand, Cisco’s Vijoy Pandey called for increased agentic collaboration across companies, calling this current movement the “Internet of Agents.”

3️⃣ Security is the differentiator

Live from our Confidential Computing Summit in June, the message was clear: security is table stakes, and it will make or break the future of AI.

Some of the biggest names in the space agree with our position:

✔️ Microsoft’s Mark Russinovich pointed to improved verifiability, calling for transparency in everything LLMs build, touch, and decide, from keyboard to production.

✔️ Galileo’s Atin Sanyal reminded us that “garbage” data sinks even the strongest models, and that using the best data, safety and security is paramount.

P.S. Any guests you’d like to see – or topics you’d like us to explore – in future episodes? Reply to this email and let us know.

Keeping it Confidential

What percent of companies report “moderately severe or worse” damages caused by AI security incidents?

1. 17%

2. 38%

3. 64%

4. 72%

See the answer at the bottom.

Code for Thought

Important AI news in <2 minutes

💤 48% of IT decision makers are losing sleep over fears that AI deployments will stop working due to unreliable support frameworks, a new Temporal.io report found. 

⚠️ 95% of executives have faced at least one problematic incident with enterprise AI, a recent Infosys report revealed, with 72% classifying the damages as "moderately severe.” 

🕵️ Recent uptick in LLM-fueled attacks from cybercriminals, spies, and researchers touted as the new era of AI hacking, according to NBC.

📄 Using DeepSeek on the job is equivalent to, “Printing out and handing over your confidential information,” a security expert from IT Pro reported. 

🛡️ Despite risks, security chiefs are deploying AI agents across cyber protocols to analyze alerts and write reports, saving time on manual tasks, the Wall Street Journal reported.

Community Roundup

Updates involving OPAQUE and our partners

⚡ The Agentic Era has arrived, and so has the CC Summit 2025 Report.

At the 2025 Confidential Computing Summit™, nearly 600 leaders from across tech and business met in San Francisco with one mission: build trust at machine speed.

The new CC Summit 2025 Report distills insights from 70+ speakers and 100+ sessions into a playbook for what’s next. 

It delves into:

• Why verifiable trust—not policies on paper—is now the baseline for enterprise AI

• How Microsoft, AMD, and CrewAI are already running confidential systems at scale

• What really holds today’s AI back

• Why security is no longer a checkbox, but the ultimate competitive edge

• The sectors already proving ROI—from finance and healthcare to manufacturing

This isn't just hype. It’s proof in production—and a roadmap for the agentic era.

📥 Download the report and give it a read.

Fresh press

😳 MIT just found that 95% of enterprise AI pilots don’t generate revenue—and the problem lies with the data.

In a report of over 300 public corporate AI initiatives, only 5% of industry leaders said that their current AI pilots were earning them additional revenue.

Poor resource allocation and a “learning gap” are listed as possible causes. But our team has a different perspective:

Without grounding your AI pilot with real business context, AI is just a demo.

With the right data, AI becomes ROI—but only if the data is properly secured.

Give the full report a read to learn more.

Open source spotlight 

🖼️ Alibaba unveiled Qwen-Image-Edit, an open-source model that swaps backgrounds, adds or removes details, and changes the orientation of images using text-based prompts.

⚡StepFunAI, a Chinese startup, launched NextStep-1, an autoregressive image generation model that produces high-quality images faster with less processing power than diffusion models. 

🎥 Skywork’s Matrix-Game 2.0 is an interactive world model that generates long, high-quality videos, offering frame-level control to users.

🤖 Jan-v1 is an agentic model, built on the Qwen3-4b-thinking model, designed with enhanced reasoning capabilities to complete complex tasks.

Quotable

🤝 “Most teams think they’re modern. Most teams also admit they’re stuck. Somewhere between post-build (“We built it!”) and pre-scale (“Oh God, how do we scale this thing?”) the real challenge begins.”

 — Temporal.io

Trivia answer: 72%

In a recent report from Infosys, 72% of executives reported having “moderately severe” damages caused by AI security incidents like privacy breaches or systems failures. Shockingly, an additional 39% reported the damage as much higher. Clearly, the need for secure AI is getting stronger.

Stay confidential!

- Your friends at OPAQUE

ICYMI: Links to past issues

• Agents are the New API Client

• 3 New Ways to Tackle AI Security

• On, In, & Transparent: A New AI Framework