😺 🎙️ PODCAST: "First major AI incident this year" (CEO warns)

Welcome, humans.

Here’s a question for ya: If China's newest open AI models are 9x to 35x cheaper than GPT-5 and outperform GPT-4.1 on coding, why shouldn't you switch immediately?

Consider the following: while everyone was distracted by OpenAI's GPT-5 launch drama, China quietly dropped two game-changers:

• Z.ai’s GLM 4.5: Beats GPT-4.1 on coding (64.2% vs 48.6% on SWE-bench), designed to be “agent-native” for autonomous task-chaining, and ~35x cheaper than GPT-5 ($0.11/$0.28 per million tokens).

• DeepSeek V3.1: a new “hybrid thinking” model with tool-calling (for agent stuff), MIT licensed, and ~9x cheaper than GPT-5 (as low as $0.07/$1.10 per million tokens).

So why wouldn't you cancel that eye-watering enterprise and make the switch ASAP?!

Because these aren't just chatbots anymore: they're models powering autonomous agents that will be interacting with your customers, or other agents across other companies.

And as we learned with earlier DeepSeek models, crushing benchmarks and rock-bottom prices doesn’t mean they’re crushing security tests.

As Navrina Singh, CEO of Credo AI, warns in our latest episode about this exact scenario: “My agent could totally poison [your] agent to do certain things that might not be aligned with the goals you've established."

Imagine that cheap, unsecured agent having access to your company's systems—then talking to your partners' agents, your vendors' agents, your customers' agents... poisoning them all to get the best outcome. This is why AI trust (and AI governance) matters.

Remember: It’s not that you CAN’T trust them… but you need to understand how trustworthy they really are before you use them.

Navrina breaks down why traditional benchmark scores aren’t telling you the full truth—and reveals what's really keeping Fortune 500 companies up at night about AI…

The spiciest revelations:

• (46:04) “I think we're going to see our first major AI incident this year” - Navrina predicts the next 6 months are critical.

• (25:40) Your AI chatbot might be recommending competitors' products - A real case from Credo AI's clients.

• (32:31) Agent-to-agent poisoning is already possible: “My agent could totally poison Neuron's agent to do certain things.”

• (28:14) China's GLM 4.5 is the first truly agentic open-source model - with “keys to building really highly performant agents.”

• (42:43) Replicate's agent just deleted someone's entire codebase - the wake-up call nobody's talking about.

Key moments you can't miss:

• (8:41) Why DeepSeek is enterprise kryptonite despite amazing benchmarks

• (13:11) The White House AI Action Plan's "moonshot moment" explained

• (35:03) “It's not what could go wrong, but what can go right if we govern AI”

• (54:02) Speaking of open models… we test OpenAI's brand new open-source model for the first time - and the prompt was basically quantum physics meets Shakespeare.

Bottom line: Companies choosing AI models based on benchmarks alone are playing Russian roulette with their business. Navrina reveals the 600+ risks Credo tracks that benchmarks completely miss—and why “trust is going to be the only moat” for AI companies.

Listen and Watch now on YouTube or Spotify.

Dive deeper with these resources:

• Credo AI's governance resources and monthly newsletter.

• The White House AI Action Plan.

• Navrina's analysis of the AI Action Plan.

• Model Trust Scores - the enterprise alternative to benchmarks.

• Learn more about Credo AI’s Model Trust Scores and why they matter. 

• Credo AI launches advisory services for enterprise AI governance.

• Major tech companies join Credo AI's partner program.

Stay curious, The Neuron Team

P.S. At 33:48, Navrina describes how your personal AI agent with access to "all your financial accounts" could get poisoned and sell your data. Sweet dreams! 😅

Stay curious,
The Neuron Team

P.P.S: Love the newsletter, but don’t want to get these podcast announcements? Adjust your preferences here.