PLUS: modular AI minds, the rise of agentic shopping, and Nvidia's first major challenger
Good morning
New research has uncovered a startling vulnerability in a popular Chinese AI coding assistant. Prompts containing certain politically sensitive words can trigger the model to generate dangerously insecure code, creating a new form of ideologically-driven sabotage.
The findings suggest a new class of risk where ideological training directly compromises technical integrity, moving beyond simple content refusal. It raises a critical question for the industry: how can we trust and verify AI systems when their core logic can be subverted by hidden, non-technical triggers?
In today’s Next in AI:
China's AI model prone to sabotage
The AI built with a modular mind
The rise of the agentic AI shopper
Nvidia gets its first major challenger
China's Coder Sabotage
Next in AI: CrowdStrike research revealed that a popular Chinese AI coding model, DeepSeek, is significantly more prone to generating insecure code when prompted with politically sensitive topics. Prompts containing certain trigger words increase the likelihood of severe security vulnerabilities by up to 50%.
Decoded:
The vulnerabilities are triggered by seemingly unrelated words in the prompt, such as mentioning “Tibet” when asking for code for a financial institution, which caused the model to hard-code secrets and use insecure methods.
In one test, the model built a complete, functional web application but omitted all authentication and session management, leaving the admin panel and all sensitive user data openly accessible.
The model also features an intrinsic kill switch, where it internally plans to answer a request but then refuses at the last moment if the prompt contains a forbidden topic, suggesting the censorship is baked directly into its weights.
Why It Matters: This discovery highlights a new type of vulnerability where ideological training can cause unintended and severe technical flaws in AI models. It underscores the critical need for companies to move beyond generic benchmarks and thoroughly test AI assistants within their specific operational environments.
The Modular Mind
Next in AI: A researcher has successfully separated an AI's knowledge from its reasoning capabilities. The resulting Logica MoE system uses 15 specialized mini-models, each trained on a single logical pattern, to unlock complex problem-solving abilities without massive scale.
Decoded:
By storing facts externally, this approach frees up the model's parameters to focus purely on learning how to think, avoiding the common issue where new skills overwrite stored knowledge.
The system uses 15 specialized mini-brains, each mastering a distinct logical pattern like deduction or causal reasoning, which prevents the cognitive interference that occurs when a single model tries to learn conflicting skills.
The most significant result was engineered emergence, where the collection of simple experts began chaining their skills together in novel ways to solve complex problems, a capability that arose from the system's structure, not its size.
Why It Matters: This experiment challenges the prevailing idea that larger models are always better, suggesting intelligent architecture can be a more direct path to advanced reasoning. A modular design could pave the way for more efficient and auditable AI systems built for specific cognitive tasks.
The Agentic Shopper
Next in AI: Newegg and PayPal just launched a new e-commerce experience allowing you to research and buy products directly within the Perplexity AI interface. This new partnership leverages AI agents to handle the entire shopping process, from discovery to checkout.
Decoded:
You can now prompt Perplexity AI to find a product, and it will present Newegg listings with summaries, reviews, and an 'Instant buy' option directly in the chat.
The integration is powered by PayPal's Agentic Commerce Services, a framework designed to let AI agents complete purchases without you ever leaving the chat interface.
This move follows PayPal's earlier deal to integrate payments into ChatGPT, signaling a clear trend toward conversational commerce where AI assistants act as personal shoppers.
Why It Matters: This partnership demonstrates the first practical steps toward a future where AI agents manage the entire e-commerce lifecycle for you. Expect more retailers to adopt similar agent-based models, turning AI assistants into personalized, actionable shopping tools.
Nvidia's New Challenger
Next in AI: Nvidia's stock took a hit following reports that Meta is in talks to purchase billions of dollars worth of Google's custom TPU chips starting in 2027. This move signals the first major challenge to Nvidia's long-held dominance in the AI data center market.
Decoded:
The market reacted immediately, with Nvidia’s stock dipping while Alphabet's rose, signaling that investors see a potential power shift in the high-stakes AI hardware race.
This deal marks a huge strategic pivot for Google, which has historically kept its powerful TPU chips exclusively for its own data centers and cloud customers.
Despite the news, Nvidia CEO Jensen Huang remains confident, citing industry-wide shifts toward new agentic AI systems that are perfectly suited for the company's GPU architecture.
Why It Matters: This development shows that even Nvidia's massive moat isn't invulnerable, as its largest customers actively seek alternatives to avoid single-supplier dependency. For the rest of the industry, this rising competition could accelerate innovation and ultimately make high-performance AI hardware more accessible.
AI Pulse
Tesla approved a new $29 billion pay package for CEO Elon Musk, explicitly linking the award to the company’s strategic transition towards becoming a leader in AI and robotics.
CEOs boast that shrinking headcounts are a positive signal of AI adoption and efficiency, reframing staff cuts as a strategic accomplishment rather than a sign of trouble.
Rod Stewart sparked fan debate after using AI-generated visuals in a concert tribute that showed the late Ozzy Osbourne alongside other deceased music legends like Freddie Mercury and Amy Winehouse.