Are LLMs Dead?

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

Today we’re exploring:

• The slowdown of LLM’s overarching power

• The rise of “lies-in-the-loop” AI cyberattacks

• Emerging open source projects worth checking out

Also mentioned in this issue: Aaron Fulkerson, Mark Hinkle, Julie Sweet, OpenAI, ABBYY, Anthropic, Microsoft, AMD, CrewAI, SF Tech Week by a16z, SkyFire, Google DeepMind, Accenture, EmbeddingGemma, Code2MCP, and K2-Think.

Let’s dive in!

This week on the AI Confidential Podcast, we’re tackling the question on everyone’s minds:

Are LLMs dead?

Spoiler alert: they aren’t, but they are slowing down.

When GPT-3/GPT-4 hit the scene, they felt like a moonshot. 

But GPT-5? It’s more of a meh update. 

It looks like we’ve officially hit the AI data plateau—and we’re quickly approaching a slide into the trough of disillusionment.

It’s the same in the enterprise space, too. 

MIT found that 95% of generative AI pilots deliver zero ROI—amounting to billions spent for little to no benefit. 

These unrealized gains reveal an uncomfortable truth: 

Public internet data is tapped out. General LLMs have hit a wall, and for enterprises to truly use genAI to its full potential, we need better inputs.

The next leap comes from grounding models in proprietary enterprise information—the real workflows, customer records, and “secret sauce” that give companies their competitive edge. 

There’s just one problem: nobody wants to risk their sensitive data leaking.

But that’s where confidential AI comes in. 

As the only solution to the slow degradation of LLMs, confidential AI gives enterprises verifiable guarantees that their data stays private, allowing teams to finally unlock insights by using AI on the information that matters most. 

It’s the way generative AI will climb out of the trough and start skyrocketing again—and at OPAQUE, our team is leading the charge.

Listen to the latest episode of our podcast, as cohosts Aaron and Mark dive deeper into this complex issue.

Check out the full episode now, you won’t want to miss it!

Keeping it Confidential

How many weekly users does ChatGPT have?

1. 165 million

2. 285 million

3. 535 million

4. 700 million

See the answer at the bottom.

Code for Thought

Important AI news in <2 minutes

🔒 More than 1 in 5 senior managers say employees are misusing their organizations’ genAI tools for personal tasks that pose security risks, a new ABBYY report found. 

🤔 Despite excitement over AI’s gains, 28% of enterprise leaders can’t figure out how to integrate genAI into their business processes, the same report found. 

🌍  Anthropic’s latest report on Claude users found that higher rates of AI adoption correlate with wealthier nations, highlighting uneven global usage.

⚔️ Researchers developed a cyberattack called “lies-in-the-loop,” which induces bad behavior in AI agents, highlighting a potential risk to software supply chains. 

🎭 North Korean and Chinese hackers are using AI to create fake military IDs and resumes as part of new and improved phishing scams.

Community Roundup

Updates involving OPAQUE and our partners

The agentic era is here—and we are stoked for the possibilities autonomous AI can bring to the table. 

If you're at the beginning of your adoption journey, our CC Summit 2025 Report is a must-read. 

This past June, nearly 600 executives, builders, and researchers converged on San Francisco for the Confidential Computing Summit™. Their shared mission: figure out how to scale trust as fast as machines move.

The takeaways are now packed into one report, drawn from 145 speakers and 100+ sessions:

• Why paper policies don’t cut it andverifiable trust is essential for enterprise AI

• How Microsoft, AMD, and CrewAI are already deploying confidential systems

• The hidden friction points slowing AI down

• Why security is the new arms race—and the sharpest competitive edge

• Where ROI is already showing up, from finance to healthcare to manufacturing

This isn’t a highlight reel. It’s a field guide for what’s working today—and where AI goes next.

📥 Download the report and give it a read.

OPAQUE events

The countdown is on: San Francisco Tech Week kicks off in just a few weeks! 🎉 

From October 6 to 12, Silicon Valley will be buzzing with founders, investors, and operators chasing what’s next in AI.

We’ll be right in the mix—hosting two exclusive gatherings on Tuesday, October 7th:

1️⃣ Agentic AI for the Enterprise: What’s the Catch?
⏰ 11:30 AM – 1:30 PM
AI agents are making headlines. But can they deliver? Leaders from Google DeepMind and Skyfire will dig into the promises and pitfalls shaping how enterprises actually use agents today.

2️⃣ AI Confidential Dinner
⏰ 6:00 – 8:00 PM
A curated dinner for the builders and decision-makers shaping applied AI. No fluff, no filler—just real conversations, sharp insights, and a chance to connect with the people pushing the field forward.

💌 Spots are going fast. Lock in your seat today.

ICYMI

Do you work in insurance? If so, check this out!

Our team recently held a webinar exploring the benefits of confidential AI in the insurance sector, and it’s filled to the brim with use cases and benefits.

Watch the webinar recording to learn:

• Use cases like workers' comp fraud detection & C-suite analytics

• How to deploy AI 4-5x faster with 60-70% lower costs

• Remote attestation to verify models, agents, and runtime

• Policy guardrails and data governance for regulated use cases

• Cryptographic proof and verifiable audit trails to satisfy compliance

• Where OPAQUE fits with platforms you already use

• A live demo of encrypted AI workflows with verifiable trust

Click here to watch the webinar recording now.

Open source spotlight 

💻 Google’s EmbeddingGemma is an open-sourced model for developers building genAI applications with RAG pipelines directly on their personal devices. 

🧩 Researchers unveiled Code2MCP, a multi-agent framework that turns GitHub code into Model Context Protocol (MCP) services, allowing agents to perform tasks on different platforms.

🇦🇪 The UAE released K2-Think, a small sovereign AI model with advanced reasoning capabilities, rivaling the likes of OpenAI and DeepSeek.

Quotable

🤖 "AI, at the enterprise level, is hard…I am talking to CEOs almost every day. Their frustration is mostly about: How do I move my organization fast enough? They recognize it's less about the technology, and more about the willingness to truly reinvent the work, the workforce."

— Julie Sweet, CEO of Accenture

Trivia answer: 700 million

According to a new paper from OpenAI, over 700 million people use ChatGPT, sending upwards of 18 billion messages a week. This level of usage is unprecedented for other technologies, representing 10% of the global adult population.

Stay confidential!

- Your friends at OPAQUE

ICYMI: Links to past issues

• Days to Seconds: Harnessing Confidential AI Agents

• What AI’s Most Innovative Thinkers are Saying

• Agents are the New API Client