U.S. distillation alert, Meta's Muse agent, and 3 more AI moves

In brief

Five developments

1. U.S. agencies warn of industrial-scale AI distillation

• The Cybersecurity and Infrastructure Security Agency, the National Security Agency, and the Federal Bureau of Investigation issued a joint cybersecurity advisory warning that China-based AI companies are conducting systematic knowledge distillation campaigns against U.S. frontier models. 1
• The advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, alleging they extracted billions of tokens and reasoning workflows from Claude, GPT, Gemini, and Grok models through unauthorized API transfer stations and prompt injection. 1
• Chinese foreign ministry spokesperson Mao Ning dismissed the accusations on Wednesday, stating China's AI progress stems from domestic technological self-reliance. 2

2. Meta launches Muse personal AI agent on dedicated secure VMs

• Meta launched Muse, a standalone personal AI agent built on its Muse Spark 1.3 foundation model to automate routine workflows across mobile devices, WhatsApp, and the web. 3
• The agent executes inside a dedicated cloud environment called Muse Secure VM, where an isolated Sentinel agent reviews outbound network traffic, protects stored credentials, and enforces user approval before financial transactions or emails occur. 3
• Muse integrates Stripe's Link payment service with built-in buyer protections and is rolling out across the United States on iOS, Android, and web browsers, with smart-glasses integration planned later this year. 3

3. Cognition AI raises $2 billion at $48 billion valuation

• AI coding startup Cognition AI raised $2 billion in a Series E funding round at a post-money valuation of $48 billion, Reuters reported. 4
• Andreessen Horowitz and Accel led the round, with additional capital contributed by existing investors Founders Fund, General Catalyst, and Avenir. 4
• Cognition stated that annualized run-rate revenue for its Devin autonomous software-engineering platform reached nearly $900 million, up from $492 million during its May 2026 round that valued the startup at $26 billion. 4

4. Google commits €13 billion to Finland AI infrastructure and nuclear power

• Alphabet's Google announced plans to invest at least €13 billion ($15.1 billion) in Finnish AI and digital infrastructure during 2027 and 2028. 56
• The capital expansion will enlarge Google's Hamina data-center operations to power computing workloads for Gemini, Search, and Maps, while contributing an estimated €3.6 billion annually to Finland's gross domestic product during construction. 56
• Google signed a 22-year power purchase agreement with Finnish utility Fortum for up to 50% of the electricity output from the Loviisa nuclear plant, marking Google's first nuclear power contract outside the United States. 5

5. Procedural Graphs introduce self-evolving execution for LLM agents

• Researchers published an architecture titled Procedural Graphs (arXiv:2609.09153), organizing agent execution into structured procedure-relation triplets rather than relying on unstructured generation over long token histories. 7
• The framework localizes an agent's active node within an execution graph, allowing a guidance model to provide step-level constraints that prevent goal drift and out-of-order tool calls during extended tasks. 7
• In benchmark evaluations across multiple reasoning datasets, a self-evolving loop using an LLM refiner edited graph topology based on validation trajectories, outperforming traditional memory baselines without manual prompt engineering. 7

The read-through

Watch next

• U.S. and China: bilateral AI safety dialogue scheduled for mid-September 2026.
• Meta: independent security assessments of Muse Secure VM and customer reception of the Sentinel approval layer.
• Cognition AI: enterprise deployment metrics for Devin following its $48 billion valuation round.
• Google and Fortum: regulatory filings and execution timelines for the 22-year Loviisa nuclear power agreement.
• Agent research: multi-lab evaluations of Procedural Graphs against traditional agent memory architectures.