AI News – April 12, 2026: The Claude Mythos Mystery, OpenAI's "New Deal," and Google's Fails

The week from April 6th to 12th, 2026, shook the Artificial Intelligence ecosystem, revealing the limits and the vertigo of the sector. The sensational leak of

If in recent months we have witnessed the consolidation of hardware and the explosion of Vertical AI, the first week of April 2026 forcefully puts frontier models and their social implications back at the center of the debate.

It was the week of Claude Mythos, the Anthropic model revealed by a data leak and deemed "too powerful" for immediate public release. While Anthropic shifts its focus to autonomous cybersecurity with Project Glasswing, OpenAI sheds its role as a tech company to don that of a legislator, proposing a veritable "AI New Deal" to the US government. Meanwhile, Google continues to battle the hallucinations of its search engine, and Nvidia reminds us that the true future of AI has a body of steel.

Here are the 5 key news stories of the week, analyzed to understand their real impact.

1. The Claude Mythos Leak: Anthropic's "Too Powerful" Model

Secrecy surrounding next-generation (AGI-level) models was shattered by a massive news leak, forcing Anthropic into an unprecedented admission.

🔍 What happened: The week opened with a sensational leak revealing the existence of "Claude Mythos", Anthropic's secret model. A few hours later, as confirmed in an official preview released to TechCrunch, the company had to move up its timeline, showcasing the system's capabilities. The strongest statement, also picked up by CBS news, is that Mythos was judged "too powerful for public release" due to autonomous reasoning and code manipulation capabilities that surpass current safety barriers (guardrails).

💡 Why it matters: This is the first time a leading company has voluntarily blocked (or at least delayed) the commercial release of its flagship product, admitting it cannot fully control its emergent capabilities. Mythos doesn't just answer complex prompts; it autonomously designs entire software systems, executing, testing, and correcting them without any human intervention.

🎯 Our take: There's a fine line between responsible safety and scarcity marketing. Saying "we've created something too dangerous for you" is the most powerful way to attract the attention of governments and military investors. However, the event marks the official start of the race for "AGI safety" as a veritable commercial product.

2. Project Glasswing: AI Becomes an Autonomous Hacker

In perfect continuity with the concerns raised by Mythos, Anthropic demonstrates why utmost caution is needed when having AI agents interact with critical infrastructure.

🔍 What happened: Concurrently with the revelations about the new models, Anthropic launched Project Glasswing, an initiative documented by The Verge. It is a cybersecurity framework where autonomous AI agents are unleashed to explore vulnerabilities within any operating system (OS) and browser. The AI performs penetration tests on a massive scale, not just reading code, but acting as a true "White Hat Hacker."

💡 Why it matters: Until yesterday, AI was used to write malicious or defensive code under human guidance. With Glasswing, the AI is the actor: it actively seeks "zero-day" vulnerabilities in servers and cloud systems before cybercriminals do.

🎯 Our take: Cybersecurity has become the definitive battlefield for Agentic AI. Whoever controls the models capable of finding vulnerabilities faster holds an immense strategic advantage, not just at the corporate level, but at the national security level.

3. Google AI Overviews: The Reliability Crisis

While theoretical power explodes, the daily, large-scale application of generative AI continues to show worrying structural limits.

🔍 What happened: A scathing technical analysis published by Ars Technica quantified the reputational damage of Google's AI-powered search engine (AI Overviews). The report found that, on complex or technical queries, the AI-generated answers from Google contain factual errors, hallucinations, or non-existent sources 10% of the time.

💡 Why it matters: A 10% error rate may seem low in a lab, but applied to the billions of daily searches handled by Google (ranging from medical advice to tax regulations), it means millions of pieces of incorrect information distributed every day. User trust in the online search monopoly is crumbling in favor of vertical, verifiable solutions.

🎯 Our take: This event reignites the conflict between "Trust vs Scale." Large Language Models are excellent probabilistic generators, but terrible search engines. As long as GenAI is used to retrieve exact facts rather than process text, hallucinations will remain an unsolvable problem.

4. OpenAI's "AI New Deal": From Technology to Geopolitics

OpenAI's transition from a research lab to a global political power (a para-state entity) has undergone a dramatic acceleration.

🔍 What happened: As reported in an analysis of the new company policies, OpenAI proposed a veritable socio-economic plan to the US government, dubbed the "AI New Deal." The document doesn't talk about parameters or tokens, but about robot taxes, sovereign wealth funds to redistribute wealth (Wealth Fund), universal basic income, and massive infrastructure investments (nuclear and data centers) to maintain American dominance over Europe and China.

💡 Why it matters: Sam Altman and company are openly admitting that the impact of their products will cause massive structural unemployment. Instead of suffering future regulation, OpenAI is trying to write it themselves, dictating the federal government's economic agenda.

🎯 Our take: The Big AI Tech companies are no longer simple software companies. They are assuming political weight comparable to that of oil multinationals in the twentieth century, attempting to shape the future social contract to protect their immense profit margins.

5. Robotics Week and "Hard Tech" Start-ups

The competition definitively shifts from chips and the cloud to the physical world: matter, mechanics, and materials.

🔍 What happened: During National Robotics Week 2026, Nvidia focused all its narrative efforts on "Physical AI": algorithms integrated into autonomous drones, industrial arms, and humanoid robots capable of learning by imitation in physical spaces. Simultaneously, MIT announced funding for 16 new companies through the START.nano program, dedicated exclusively to "Hard Tech": new materials for semiconductors, bio-integrated sensors, and nanotechnologies to support the next wave of AI computing.

💡 Why it matters: There is saturation in the world of purely generative software. The market has understood that the true trillion-dollar business is applying the "brain" of AI to robotic "bodies" and advanced infrastructure. Without new milestones in materials science (to dissipate heat and reduce data center consumption), AI development would stall.

🎯 Our take: As highlighted in our anniversary editorial, the horizon is no longer generating nice text, but making a machine move intelligently on a construction site or in an operating room. It is the triumph of "Embodied" AI.

FAQ: Frequently Asked Questions of the Week

1. What is Claude Mythos and why was its release blocked? Claude Mythos is the codename for the next-generation model (presumably the successor to the Claude 3.5/4 family) developed by Anthropic. A leak revealed its existence and incredible logical capabilities. Anthropic stated that the model possesses "autonomous agency" abilities (writing, correcting, and executing code without supervision) so advanced that they represent a risk to global cybersecurity if released open source or without adequate "guardrails."

2. What does "Project Glasswing" mean for cybersecurity? It is a project that uses AI not for passive defense, but for proactive attack. The AI scans millions of lines of code in operating systems (like Windows, Linux, or iOS) and web browsers to uncover vulnerabilities invisible to human programmers. This allows companies to "close the doors" before cybercriminals can exploit those same flaws.

3. Why does Google AI Overviews get 10% of answers wrong? The models underlying AI Overviews (like Gemini) are Large Language Models (LLMs). They are designed to predict which word makes the most statistical sense after the previous one, not to query a database of "true facts." When they encounter conflicting sources on the web or lack sufficient data, they tend to "invent" an answer just to satisfy the user's query (the hallucination phenomenon), making them dangerously unreliable on complex technical or scientific topics.

4. What is the "Robot Tax" proposed by OpenAI? It is an economic concept whereby if a company replaces a human worker with Artificial Intelligence (or a robot), it must pay a tax equivalent to the income taxes that human worker would have paid to the state. OpenAI proposes that the proceeds from these taxes, combined with sovereign wealth funds (Wealth Funds), be used to fund a universal basic income for people who will lose their jobs due to automation.

5. What is meant by "Hard Tech" and "Physical AI"? Physical AI is artificial intelligence integrated into robots, drones, and machinery, capable of perceiving the physical environment through sensors and performing mechanical actions. Hard Tech (or Deep Tech) refers to science-intensive innovations based on tangible materials, such as new materials, nanotechnologies, or quantum computing. These are the physical and manufacturing sectors that must evolve to support the future AI infrastructure.