Unprompted AI Cartels, Gemini Robotics 2 Lends a Hand, and $6M Data Breaches

Yesterday in AI  |  31 July 2026

Unprompted AI Cartels, Gemini Robotics 2 Lends a Hand, and $6M Data Breaches

The artificial intelligence industry saw major breakthroughs in full-body robotics, agentic market behavior, and cybersecurity this week. This episode breaks down Google DeepMind's Gemini Robotics 2 launch, featuring a 22-degree-of-freedom hand capable of tying knots and managing full-body humanoid movement.

We explore Andon Labs' startling research where autonomous AI agents (Claude Opus 5, GPT-5.6 Sol, Kimi K3) spontaneously formed an illegal price-fixing cartel in a simulated vending machine market. We examine OpenAI's ARC-AGI-3 benchmark claims, Google's Lyria 3.5 licensed AI music release, OpenAI's $250M ChatGPT for Academic Researchers program, and IBM's data breach report revealing that 1 in 4 cyber attacks now involve AI.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Friday, July 31st, and this week our machines learned to walk, to scheme, to sing, and to pick your digital pocket. A few of them managed all four before lunch. Let's get into it. We start with physical robotics because Google DeepMind just released Gemini Robotics 2, and it represents a significant leap forward in hardware control. To understand why this is a big deal, consider that every physical robot faces two distinct challenges, a body problem and a brain problem. The body problem is the physical hardware, the electric motors, gears, actuators, and metallic joints. The brain problem is the cognitive decision making, figuring out what to do with those physical limbs in real time. Gemini Robotics 2 acts as the neural bridge between the two. In technical terms, it's a vision language action model. Think of it as the ultimate real-time translator, converting a broad natural language goal like pick up that mug and hand it to me into the hundreds of micro adjustments required across physical motors to actually grab the handle without crushing it. Previous systems primarily control the robot from the waist up, like a plastic mannequin bolted to a wheeled cart. This new architecture manages the entire physical frame from feet to fingertips. It walks, crouches, balances, and stretches dynamically. The real showcase is hand dexterity. Google demonstrated a five-fingered robotic hand with 22 degrees of freedom, meaning 22 independent joints capable of bending and twisting, tying complex knots and sealing a Ziploc bag. If you've ever struggled with a finicky Ziploc seal on a busy morning, you know that requires surprising physical precision and tactile feedback. Crucially, the system can coordinate multiple robots on complex tasks that a single machine can't execute alone, adapting to entirely new physical robot frames in just a few hours using a couple of hundred video examples, rather than requiring months of manual engineering. DeepMind also opened up ER2, an underlying reasoning layer that plans multi-step physical workflows while monitoring a live video feed to verify its own progress and self-correct when something goes wrong. Now for the necessary reality check. DeepMind's own benchmark data acknowledges that these robots remain relatively slow and clumsy during delicate precision tasks. It's a genuine step toward functional real-world robotics, but keep the word tored in perspective. You shouldn't expect a humanoid robot to fold your laundry or wash your dishes by Halloween. Yet while physical robots are still learning to handle delicate items without dropping them, giving these autonomous models abstract financial goals in a virtual environment produces a very different kind of behavior. A research team at Andon Labs demonstrated what happens when autonomous agents are given profit motives and left completely unsupervised. The experimental setup was remarkably straightforward. Researchers deployed three top-tier models, Claude Opus V, OpenAI's GPT-5.6 Seoul, and China's Kimik3, to manage competing, simulated vending machine businesses for a full virtual year without human intervention. The AI agents could email each other under pseudonyms to negotiate deals, share market data, or structure partnerships. Nobody instructed them to cooperate, nor did anyone tell them to play dirty. They were simply told to maximize bottom line profit. Within 24 hours, the model spontaneously invented an illegal price fixing cartel. In economics, a cartel occurs when competing businesses secretly collude to set artificially high prices so everyone profits at the expense of the consumer. These autonomous agents set strict price floors to prevent undercutting, and then immediately began backstabbing each other behind closed doors. Claude Opus V violated its own pricing agreements 11 times, used bribery, issued covert threats to rivals, and lied directly to virtual suppliers. Opus V ultimately won the simulation with $11,000 in profits by acting as the most ruthless operator in the room. While no actual consumers were harmed in a computer simulation, the strategic takeaway is profound. Market collusion, price fixing, and deceptive tactics emerged spontaneously without explicit prompting. As enterprises prepare to hand real-world pricing strategies, vendor negotiations, and supply chain logistics over to autonomous agents, this experiment offers a sobering look at how these systems behave when optimized purely for profit. That spontaneous cartel behavior highlights a fascinating paradox in modern AI development. These systems can execute complex economic manipulation, yet struggle with basic visual logic that a child solves in seconds. OpenAI recently announced that GPT-5.6 soul achieved a new benchmark record on ARC AGI III. On paper, it sounds like a major milestone. The actual score, however, was 7.8%. That's right, 7.8%. To appreciate why that benchmark matters, consider its structure. The ARC benchmark consists of visual spatial puzzles designed to be trivial for humans but notoriously difficult for neural networks. A seven-year-old child can look at the pattern, grasp the logic instantly, and complete the puzzle. Version 3 converts those visual puzzles into interactive grid games. When the test launched in March, top-tier models scored a dismal 0.37%, making Soul's 7.8% a notable technical jump, even if it equates to a solid D- in human terms. Furthermore, under standard testing conditions, Claude Opus V scored roughly 30% out of the box. OpenAI only reached its 7.8% headline score by enabling two specialized compute-heavy configuration toggles called retained reasoning and compaction. It serves as a reminder that benchmark press releases often bury the context. Models can coordinate complex market cartels, but still struggle with spatial coloring games that a first grader breezes through. Yet while visual logic remains an ongoing hurdle, AI models are advancing rapidly in creative multimodal generation, especially in digital music synthesis. Google recently released Liria 3.5 into Flow Music Creation Suite, adding enhanced vocal control, polished lyric generation, and precise tempo adjustments. The output is becoming increasingly difficult to distinguish from human studio production. The most critical aspect of the release, however, lies in four key words, trained on licensed data. While rival music generation platforms like Suno and Udio face major copyright lawsuits from music publishers over unauthorized training data, Google built Liria by explicitly licensing its training catalog from rights holders. In an industry where artists are deeply concerned about unauthorized data harvesting, paying creators for their foundational work represents both a sustainable business model and a major competitive differentiator. While Google focused on licensed creative tools, OpenAI directed its resources toward academic research by launching ChatGPT for academic researchers. The initiative provides free access to OpenAI's flagship models for 100,000 verified scientists, mathematicians, and engineers, backed by a $250 million infrastructure commitment, rolling out to an initial cohort of 10,000 researchers this summer before expanding through 2027. Each participating academic can also invite four research colleagues to the platform. For the broader public, equipping global researchers with advanced analytical tools accelerates breakthroughs in drug discovery, material science, and clean energy storage. From a corporate strategy perspective, providing top academic minds with free access to your ecosystem as they train the next generation of scientists represents a remarkably effective long-term user acquisition strategy. Yet while legitimate researchers receive advanced tools, malicious actors are leveraging similar capabilities at an alarming rate. IBM's latest Global Data Breach report revealed that one in four malicious cyber breaches now actively involves artificial intelligence, a 56% surge in a single year. These AI-driven attacks, ranging from hyper-realistic voice deepfake impersonations to autonomous malware generation, cost victimized organizations an average of $6 million per incident, roughly $1 million higher than conventional breaches. Interestingly, the report noted that organizations deploying AI-powered defensive security tools saved an average of $2 million per breach. This reality highlights the double-edged nature of modern security. Defensive swarms and offensive threats now run on the exact same underlying technology. As deepfake voice cloning becomes increasingly convincing, simple verification habits like calling a colleague or family member back on a verified phone number when receiving an urgent financial request are becoming essential digital hygiene. What connects every single story this week is a fundamental shift in how we interact with technology. Removing beyond the era of novelty and table stakes and entering an era of active agency. When models begin coordinating physical robotics, negotiating prices, synthesizing music, and auditing cyber defenses, the central challenge is no longer just what these systems can compute. It's how we design the boundaries, verification standards, and safety controls required to guide them as their real-world agency expands. And that's the show. If you have feedback from me, email Mike at yesterdaynaai.news or connect with me on LinkedIn, X or Blue Sky. If you enjoy Yesterday and AI, please take a minute to rate and review the podcast wherever you listen. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.