The artificial intelligence sector is facing a turbulent week marked by aggressive regulatory intervention and a damaging transparency crisis at a leading AI lab. In a landmark move, Chinese regulators have successfully dismantled a completed cross-border AI acquisition, forcing Meta to unwind its $2 billion purchase of the agentic AI startup Manus. Simultaneously, Anthropic is scrambling to contain backlash after secretly degrading the output of its newest flagship model, Claude Fable 5, for users suspected of building competing AI systems.
China’s National Development and Reform Commission (NDRC) ordered the reversal of the Meta–Manus deal in April 2026, citing violations of foreign investment and technology export rules. It is the first time Chinese authorities have forced a completed AI acquisition to be unwound. An internal memo at Meta instructed staff to “sunset” Manus and migrate projects onto Meta’s own platforms, while a data firewall now separates the two companies. Manus co-founders Xiao Hong and Ji Yichao were summoned to Beijing and barred from leaving the country as part of the investigation. The startup’s early research was conducted in China using Chinese talent and data, and regulators argued that relocating headquarters to Singapore was an attempt to evade oversight. The co-founders are now exploring a roughly $1 billion buyback at the same $2 billion valuation, complicated by earlier investors such as Tencent and ZhenFund.
The forced separation deepens the tit-for-tat technology cold war between Washington and Beijing. The U.S. has restricted American investment in Chinese AI, semiconductor, and quantum firms, while China has tightened rare-earth licensing, banned foreign AI chips from state data centers, and instructed private tech firms to reject American investment without explicit approval. As Brookings fellow Kyle Chan noted in Congressional testimony, China’s top AI models still lag behind American frontier models by several months or more.
On the other side of the AI industry, Anthropic endured a 48-hour firestorm after launching Claude Fable 5 – the first of its new Mythos class – with an invisible safeguard that secretly degraded responses for users flagged as potentially building competing AI models. Unlike visible filters for cybersecurity and biology that reroute users to the older Opus 4.8 model, the LLM-development classifier silently altered behavior through prompt modification or parameter tweaks, delivering contaminated results. AI research firm SemiAnalysis publicly called out the company after its GPU inference work was flagged, triggering broader outrage over reproducibility in research.
By Thursday, Anthropic apologized and announced that flagged requests will now visibly fall back to Opus 4.8 with a stated reason, admitting that “invisible safeguards… was the wrong tradeoff.” However, making safeguards transparent also makes them easier to bypass, forcing the classifier to cast a wider net and likely increasing false positives. The fix does not remove the restrictions entirely, leaving critics unsatisfied.
