AI Tech Daily March 29, 2026

Executive Summary

Today saw a convergence of significant developments shaking the foundations of the AI industry: regulation, safety, and the emergence of next-generation models. In Europe, the realistic timeline for AI regulation was recalibrated, while at the development front, Anthropic’s powerful new model was revealed. Furthermore, Google DeepMind presented a new approach to quantifying AI manipulation risks, highlighting both the risks associated with AI’s rapid proliferation and the focus on responsible development.

Today’s Highlights

1. European Parliament Approves Delay in EU AI Act Application Deadline

The European Parliament overwhelmingly approved, with 569 votes in favor and 45 against, a “simplification (omnibus)” proposal that postpones the application deadline for parts of the “EU AI Act.” This means that the regulations for high-risk AI systems will be delayed from the initially scheduled August 2026 to as late as December 2, 2027.

This decision stems from the reality that the European Commission is behind schedule in releasing detailed technical guidance for high-risk AI. It was determined that enforcing strict regulations without clear, specific criteria for companies and developers to comply would lead to market disruption. Additionally, the application of rules concerning floods and watermarking for AI-generated content has also been granted a grace period until November 2026.

This decision is a crucial step towards providing “predictability” and “legal certainty” to the industry. However, experts warn that “even with the delay, CIOs and legal officers should prepare as if the standards are already in effect.” This situation exposes a typical challenge in modern AI governance: the slow pace of building regulatory and standardization infrastructure compared to the rapid evolution of AI.

Source: European Parliament Official Press Release

2. Anthropic’s Powerful New Model “Claude Mythos” Leaks

A misconfiguration in Anthropic’s content management system led to the public leak of unreleased draft documents concerning the company’s next-generation AI model. The documents hinted at a model named “Claude Mythos” (or “Capybara”), described as delivering a “step change” in capabilities over the current Claude Opus in tasks related to cybersecurity, coding, and reasoning.

Particularly noteworthy is the model’s “cyber capability.” The internal documents state it possesses significantly more powerful vulnerability discovery and exploitation abilities than any current AI model, with descriptions suggesting a potential advantage for attackers over defenders. In response, Anthropic has stated its plan for a very cautious and security-focused, limited release.

This incident serves as a stark reminder of the importance of assets held by leading AI companies and demonstrates that the “dual-use” risk of AI is no longer theoretical but is being incorporated into actual development plans. Investors and the cybersecurity industry are increasing their calls for transparent disclosure from Anthropic, and the progress of future hands-on testing is keenly awaited.

Source: The Decoder News

Other News

• Google DeepMind Releases Framework for Evaluating Harmful AI Manipulation Risks Google DeepMind has released its first empirical toolkit for measuring the potential for AI models to psychologically manipulate or influence humans. Based on research involving over 10,000 participants across the UK, US, and India, it evaluates AI’s impact in high-risk environments such as health and finance. The research demonstrated that “consistency in success is lacking” when evaluating AI safety (success in one domain does not guarantee success in another). Source: Google DeepMind Blog

• Microsoft Research Releases New Benchmark “GroundedPlanBench” for Robot Manipulation Microsoft Research has introduced “GroundedPlanBench,” a benchmark for evaluating long-horizon task planning by robots. It assesses the ability to spatially ground visual information and complete physical manipulations according to complex instructions. Evaluation results using the multimodal model “Qwen3-VL” suggest that an end-to-end grounding approach may be superior to traditional methods that separate planning and grounding. Source: Microsoft Research Blog

• NVIDIA Presents New Strategy for Improving Energy Efficiency in AI Factories NVIDIA has explained a technical approach for flexibly optimizing the energy consumption of AI data centers (AI factories) in coordination with the power grid. The goal is to achieve “power-flexible” operations that reduce grid load while meeting the surging demand for computation. This is based on the outlook that infrastructure sustainability will be a decisive factor in AI competition. Source: NVIDIA Blog

• OpenAI Details Philosophical and Technical Background of “Model Spec” OpenAI has released detailed background information on the intent and implementation process behind the latest version of its “Model Spec,” which governs ChatGPT’s behavior. It visualizes discussions, particularly concerning “principles for protecting users under 18” and “what constitutes default behavior versus user-overridable settings,” aiming to enhance AI governance and transparency. Source: OpenAI Research Blog

• Searchless.ai Launches as Media Specializing in Transition from Search to AI-Mediated Discovery A new media outlet, “Searchless.ai,” has launched to track the structural shift from traditional search engines to “AI-mediated discovery,” where AI directly generates answers. It plans to comprehensively analyze the rise of a new marketing field called “Generative Engine Optimization (GEO)” in response to declining search click-through rates. Source: Newsfile Corp Release

Summary and Outlook

Today’s news indicates that the AI industry has fully transitioned from “experimentation” to “practical application.” The delay in the EU AI Act’s application deadline signifies a period of practical adjustment for social implementation, while Anthropic’s new model and Google DeepMind’s safety evaluation framework underscore the reality of rapidly expanding AI capabilities and risks.

Key points to watch moving forward include:

1. Standardization of AI Governance: Discussions will likely shift from individual corporate ethics to establishing industry-wide “unified evaluation criteria.”
2. Integration of AI and the Physical World: AI is increasingly exerting physical influence beyond the digital realm, seen in robot manipulation and AI factory power management.
3. Transformation of Digital Access: The shift in dominance from traditional search engines to AI agents requires close observation as it fundamentally rewrites web traffic models.

References

This article was automatically generated by LLM. It may contain errors.