PLUS: The simple new brain for AI agents, Google's AI coding interviews, and Apple vs OpenAI
Happy reading
Apple's flagship M5 chip security, a feature that took five years to develop, has been publicly bypassed. The exploit was created in just five days with the help of Anthropic's new Mythos AI model.
The event demonstrates a dramatic acceleration in discovering critical vulnerabilities, powered by advanced AI. With even the most robust hardware defenses now susceptible to AI-driven attacks, how will the cybersecurity landscape adapt to this new reality?
In today’s Next in AI:
Anthropic's AI cracks Apple's M5 chip
A new, simpler brain for AI agents
Google pilots AI-assisted coding interviews
The growing rift between Apple and OpenAI
AI's 'Bugmageddon' Begins
Next in AI: Security researchers, working with an early version of Anthropic's Mythos AI, developed the first public exploit for Apple's brand-new M5 chip. The breakthrough bypassed the Mac's most advanced hardware security in only five days, hinting at an impending AI “Bugmageddon.”
Explained:
The exploit defeated Apple's flagship security feature, Memory Integrity Enforcement, which is built directly into the M5 hardware to stop this exact class of attack.
While Apple invested five years developing this protection, the research team from Calif used Mythos to build a working exploit in just five days.
The AI-human team chained two separate bugs to create a privilege escalation exploit, a technique that can give an attacker full control over the computer.
Why It Matters:
This shows how advanced AI models are dramatically accelerating the discovery of critical software vulnerabilities. It signals that even the most robust hardware defenses must now be developed with AI-driven attack methods in mind.
The Agent's New Brain
Next in AI: A growing movement of developers is abandoning complex vector databases for a simpler, more powerful brain for their AI agents: plain markdown files managed with Git. This back-to-basics approach is proving more robust, maintainable, and effective for building long-term agent memory.
Explained:
This shift reframes agent memory as an organizational challenge, not just a technical one. Instead of getting lost in abstract embeddings, this approach creates a unified knowledge infrastructure that both humans and AIs can easily read, edit, and maintain, a method validated by Y Combinator CEO Garry Tan's GBrain project.
The core architecture uses Git's version control to give agents a true sense of history. You can see exactly how an agent's understanding evolves, revert mistakes, and allow multiple agents to work on the same knowledge base safely—all with human-readable storage that makes debugging simple.
The result is a system that compounds knowledge rather than just retrieving it. By using familiar tools, teams can build a maintainable knowledge base that grows smarter over time, a key principle for getting agentic AI right in a business setting.
Why It Matters: This approach makes building powerful AI agents more accessible by using tools developers already know and trust. It signals a major shift toward creating AI systems that are transparent and auditable, integrating them directly into human workflows instead of treating them as black boxes.
Google's AI Interview
Next in AI: Google is piloting a new interview format that allows software engineering candidates to use an AI assistant, signaling a major shift from testing rote memorization to evaluating real-world collaboration skills.
Explained:
The pilot program, starting in H2 2026 for junior and mid-level U.S. roles, will let candidates use Gemini during the code comprehension portion of the interview to read, debug, and optimize code.
This change aligns interviews with the reality of modern engineering, where AI now helps generate 75% of new code at Google and 80% at OpenAI.
Instead of banning AI, interviewers will now evaluate a candidate's AI fluency, focusing on practical skills like prompt engineering, output validation, and debugging AI-generated code.
Why It Matters:
This move formalizes how top engineers already work, integrating AI tools directly into their daily workflows. For tech professionals, mastering AI collaboration is no longer just a productivity hack; it's becoming a core, measurable competency for getting hired.
Apple's OpenAI Rift
Next in AI: The two-year partnership between Apple and OpenAI is reportedly fraying, with the AI startup disappointed by ChatGPT's limited integration in Apple's ecosystem and preparing for potential legal action.
Explained:
Unmet revenue goals are a primary source of tension, as OpenAI expected billions from subscriptions driven by the integration but has seen minimal financial returns.
The integration's design has proven ineffective, requiring users to explicitly invoke "ChatGPT" in Siri and providing constrained responses that push users back to the standalone app.
Tensions are also escalating as Apple plans to integrate rival models from Google and Anthropic, while OpenAI's own hardware ambitions create a new competitive front.
Why It Matters: This fallout highlights the immense challenge of integrating third-party AI into tightly controlled tech ecosystems like Apple's. The outcome will set a major precedent for how future partnerships between platform giants and AI innovators are structured.
AI Pulse
x.ai launched Grok Build, a new coding agent and command-line interface designed for professional software engineering that can plan complex tasks, delegate work to parallel subagents, and integrate with existing developer workflows.
Anthropic partnered with the Gates Foundation, committing $200M in funding, Claude credits, and technical support for programs in global health, education, and economic mobility.
Resemble AI released DramaBox, an open-weight text-to-speech model that interprets stage directions written into prompts to generate more expressive, performance-driven audio.
Ontario's Auditor General found that AI scribe systems approved for healthcare providers routinely fabricated information, inserted incorrect drug data, and missed key details from patient recordings.