In the early hours of May 24, 2026, the OpenAI Codex platform accidentally exposed several unreleased models to Pro users after mistakenly deploying an internal testing environment to production. Among the leaked models was GPT-5.5 (codenamed oai-2.1), described as the 'latest frontier agent coding model,' with performance potentially 3–4 times faster than GPT-5.4. Also revealed were Glacier-series models, whose descriptions suggest a possible architecture beyond Transformer; and the life sciences–specific Heisenberg model, signaling OpenAI's entry into protein folding and drug discovery. OpenAI President Greg Brockman later praised GPT-5.5 on social media as a 'very good model.' The leak reveals that OpenAI is advancing in parallel across multiple technical fronts—including agents, novel architectures, and life sciences—far beyond public awareness.

The White House has approved a $9 billion classified funding initiative to help intelligence agencies such as the CIA and NSA acquire advanced AI chips and computing infrastructure. Approved on May 22, the funds will be used to purchase Nvidia Grace Blackwell superchips and build dedicated data centers capable of operating in secure, classified environments. This aims to address the current inability of intelligence agencies to deploy cutting-edge AI models due to a lack of advanced semiconductors. Anthropic's Mythos model has already been integrated into NSA systems. As generative AI plays an increasingly critical role in satellite image analysis, communications interception, and open-source intelligence processing, the government is racing to close the technological gap with commercial AI labs to prevent it from becoming a national security risk.

DeepSeek announced a permanent price adjustment for its flagship model V4-Pro API, reducing it to one-quarter of the original price by making previous promotional pricing permanent. After the update, cached input costs 0.025 RMB per million tokens, uncached input costs 3 RMB, and output costs 6 RMB. This pricing makes V4-Pro approximately one-third the cost of Gemini 3.1 Pro Preview, one-twelfth of GPT-5.5, and one-nineteenth of Claude Opus 4.7. Founder Liang Wenfeng stated the goal is to operate 'without subsidies or excessive profits,' emphasizing that AI should be affordable for everyone. This move, combined with DeepSeek’s ongoing ~70 billion RMB fundraising round, reflects a strategy to rapidly expand market share through aggressive pricing while maintaining long-term AGI ambitions.

On May 22, Meituan's tech team officially open-sourced the digital human video generation model LongCat-Video-Avatar 1.5, achieving comprehensive upgrades in lip synchronization, physical plausibility, long-video stability, and inference efficiency. With the introduction of DMD distillation technology, the number of generation steps was reduced from 50 to 8, improving inference efficiency by approximately 15x—enabling 10-second video generation in about 1 minute. Evaluations show leadership over models like Kling Avatar 2.0, OmniHuman-1.5, and HeyGen in dimensions such as physical plausibility and temporal stability, with user preference win rates of 65.9%, 61.1%, and 54.3% respectively, and a low frame-skipping rate of only 0.8% in multi-person scenarios. The project has been simultaneously released on GitHub and HuggingFace.

On May 22, 2026, Japan's Financial Services Agency (FSA) and Bank of Japan (BOJ) jointly instructed financial institutions to urgently respond to cyber threats posed by frontier AI. Advanced AI models can rapidly identify and exploit software vulnerabilities, drastically compressing traditional defense timelines. Institutions are required to prioritize critical systems, accelerate patch deployment, eliminate technical debt, and predefine criteria for proactively suspending key services when necessary. This is not merely a technical issue but demands executive-level engagement across budgeting, staffing, and business continuity planning. The directive comes amid growing concerns that frontier AI models like Anthropic's Claude Mythos are shifting the balance of cybersecurity offense and defense.

Transparency obligations under Article 50 of the EU Artificial Intelligence Act will take effect on August 2, 2026, applying to roughly one-third of organizations. The provision covers four scenarios: clear disclosure when AI interacts with users, machine-readable labeling of generative AI content, notification to individuals when emotion recognition systems are used, and mandatory disclosure of deepfakes. Disclosures must be 'clear and prominent,' not buried in terms or footnotes. Violations may incur fines up to €15 million or 3% of global annual turnover. A comprehensive agreement reached on July 5, 2026, provides a transition period for AI systems already on the market before August 2, allowing machine-readable labeling requirements to be delayed until December 2.

Nous Research has introduced the Contrastive Neuron Attribution (CNA) method, which identifies key MLP neurons controlling refusal behavior in instruction-tuned models via forward pass analysis. The method requires no training or weight modification—ablating only 0.1% of activated neurons reduces refusal rates by over 50%, while maintaining output quality above 0.97 and less than a 1 percentage point change in MMLU accuracy. Validated across 16 Llama and Qwen series models (ranging from 1B to 72B parameters), a key finding is that the late-layer structures distinguishing harmful from benign inputs already exist in pre-trained base models. Alignment fine-tuning does not create new structures but repurposes existing neurons into sparse, directionally intervenable refusal gates.

The National Cybersecurity Standardization Technical Committee has released the 'Artificial Intelligence Application Ethics and Safety Guidelines 1.0,' systematically establishing ethical safety baselines for AI applications. Covering core principles such as human oversight, fairness, privacy protection, and controllability, the guidelines apply across the entire lifecycle of development, service provision, and usage. Integrating flexible guidance with binding regulation, they promote cross-departmental coordination and societal co-governance, clarify responsibilities for all parties, strengthen risk assessment, transparency, and emergency mechanisms, and emphasize protection of vulnerable groups such as minors. The guidelines also support the advancement of key technologies like trustworthy AI and privacy-preserving computation, contributing China's approach to global AI governance.