According to the Financial Times, Anthropic has dispatched approximately six engineers to the U.S. National Security Agency (NSA) to assist in deploying its cutting-edge cybersecurity AI model, Mythos. It remains unclear whether the model is already actively involved in the NSA's offensive cyber operations. Previously, the U.S. Department of Defense labeled Anthropic a 「supply chain risk」 and banned its use after the company refused to allow its technology to be used for mass surveillance and autonomous weapons. Despite this restriction, the NSA continues to utilize Mythos. Anthropic has expressed concerns about potential misuse of its cybersecurity capabilities and has restricted access accordingly. Neither party has commented. On the same day, Japan's Hitachi also announced it would adopt Mythos for infrastructure network defense, highlighting the model's accelerating penetration into both government and enterprise sectors.

SpaceX and Google have entered a 32-month AI compute leasing agreement, with Google paying $920 million per month, covering approximately 110,000 NVIDIA GPUs. The deal aims to provide transitional compute capacity for Google's Gemini Enterprise agent platform to meet unexpectedly high customer demand. SpaceX is monetizing data centers originally built for xAI's Grok model, despite its AI division recording a $2.5 billion operating loss on $818 million in revenue. This transaction marks a key step for SpaceX to demonstrate its ability to monetize AI infrastructure ahead of its IPO. Alphabet invested in SpaceX back in 2015 when its valuation was just $12 billion; SpaceX now targets an IPO valuation exceeding $1.75 trillion.

Former Salesforce Chief Scientist Richard Socher has announced the launch of Recursive Superintelligence, securing $650 million in initial funding. The company is dedicated to developing recursively self-improving AI models capable of autonomously identifying operational weaknesses and redesigning their own architecture. Its core technology is based on an 「open-ended」 framework that automates the entire cycle of research ideation, implementation, and validation. The team includes Peter Norvig, former Google DeepMind head Tim Rocktäschel, and Josh Tobin, early leader of OpenAI Codex. The team employs a 「rainbow team」 co-evolution approach, where subordinate AIs continuously test the primary model through millions of automated iterations to uncover flaws, aiming to shorten the R&D cycle to quarterly intervals.

Chinese startup Spirit AI has achieved first place on the RoboArena robotics benchmark—jointly developed by NVIDIA, Stanford, and Berkeley—with its embodied intelligence model Spirit v1.6 scoring 1924 points, surpassing NVIDIA's Cosmos3-Nano-Policy at 1881. It is the first Chinese model to top the leaderboard, overtaking NVIDIA just two days after its release. On the same day, Spirit AI announced a new financing round of 1.5 billion RMB (approximately $222 million), bringing its valuation above $1.4 billion. In 2025, venture investment in China's robotics sector reached $3.4 billion, 42% higher than in the U.S. Experts note that data is the biggest bottleneck for physical AI, and China holds structural advantages in manufacturing scenarios and data collection. In response, NVIDIA subsequently formed the Cosmos Consortium to maintain competitiveness.

The world's first AI-assisted designed vaccine, developed by the University of Oxford and Basecamp Research, has entered human clinical trials. Targeting Crimean-Congo Hemorrhagic Fever—a tick-borne, high-risk virus with a fatality rate as high as 40%, for which no widely approved human vaccine currently exists—the AI system rapidly identifies key antigen fragments from vast genomic and biological datasets that can trigger strong immune responses, significantly reducing the time and cost associated with traditional vaccine development. The first-phase trial, led by the Oxford Vaccine Group, will evaluate safety and immunogenicity. This project marks a new phase in AI-driven drug discovery and could accelerate responses to emerging infectious diseases and global health threats.

New York-based AI company Flourish has completed a $500 million funding round, reaching a $2.5 billion valuation. Jeff Bezos personally invested nearly $100 million, with participation from Alphabet's GV, Lux Capital, and Catalio Capital. Co-founded by Thomas Reardon, former lead developer of Microsoft Internet Explorer, Flourish aims to develop energy-efficient AI models that mimic biological cortical columns, targeting power consumption between 20 and 50 watts—far below the hundreds of watts consumed by current GPU servers. The technology relies on algorithmic and architectural innovation rather than specialized chips, enabling deployment on existing hardware. The company plans to build an internal neuroscience lab and expand collaboration with chip manufacturers. Despite having no product or revenue yet, the high valuation reflects strong market demand for breakthroughs in AI energy efficiency.

On June 5, Microsoft released a report identifying seven new types of AI agent attacks, expanding its previously published taxonomy of failure modes for agent-based AI systems. The newly identified attack vectors include: agent supply chain tampering, goal hijacking, inter-agent trust escalation, computer-use agent vision attacks, session context contamination, MCP/plugin abuse, and capability architecture information leakage. These findings stem from trends such as rapid AI adoption, maturation of the MCP ecosystem, and the rise of computer-using agents. Microsoft recommends security teams adjust their defenses accordingly—such as inventorying supply chains, generating software bills of materials for each deployed agent, using verifiable credentials for identity authentication, and incorporating these new threats into red team exercises and human-AI audit workflows.

Anthropic is advancing preparations for the public release of its new model Oceanus, which had previously been provided to red teams for security testing. However, due to individuals reselling access to the model via Chinese API proxies, Anthropic was forced to suspend red team evaluations, potentially delaying Oceanus' official launch. Based on Anthropic's past release patterns, red team testing typically occurs about one week before broad availability, suggesting Oceanus was already nearing public rollout. This leak highlights the supply chain control challenges faced by frontier AI models during security assessment phases and raises new questions for AI companies on protecting unreleased models—particularly regarding governance of cross-border API resale through gray-market channels.

Confirmed by CNBC, OpenAI CEO Sam Altman and the Trump administration are in discussions over the U.S. government taking an equity stake in OpenAI. The idea was initially proposed by Altman in 2025 and has gained momentum during his recent Washington visits with lawmakers. OpenAI's April policy proposal outlined a concept for a 「public wealth fund」 that would invest in long-term assets, allowing American citizens to benefit from AI-driven economic growth. This aligns with Trump's executive order calling for the creation of a sovereign wealth fund. The discussions are part of a broader set of recent White House AI initiatives, including directives on national security AI adoption and a voluntary pre-release review mechanism for frontier models. Specific equity stakes and terms have not yet been finalized.