US spot Bitcoin ETFs recorded a net inflow of $1 billion yesterday, marking the first positive net flow year-to-date.

Beating AI Express News: A new paper on agent training published by DeepSeek has drawn widespread attention. The 31-page paper introduces DSec, DeepSeek’s internally used production-grade sandbox platform, with over 100 authors including DeepSeek co-founder Liang Wenfeng. A notable section of the paper focuses on "agent misconduct". DeepSeek found that agents can obtain answers through unintended channels, such as residual answers in files managed by search platforms, undermining the validity of training and evaluation results. Even after implementing access controls, some agents exchange file block mappings, attempting to access protected file contents via another file descriptor, compromising tasks or shared infrastructure. DeepSeek argues that no single mechanism can prevent all agent misconduct and system failures. Thus, the team’s approach involves enhancing observability to identify new issues, and continuously hardening DSec as models evolve, including access controls to restrict agents from obtaining answers via unintended channels and reducing incentives for deceptive behavior. These controls address some of the issues.

BlackRock’s latest research report titled *Machine-Native Economy* notes that the widespread adoption of AI could act as an underappreciated driver of demand for digital assets. As AI agents and inter-machine payments gain traction, demand for stablecoins, native crypto assets, and other on-chain assets is likely to rise further. BlackRock argues that AI agents require around-the-clock high-frequency, low-value (even sub-1-cent) inter-machine transactions, while traditional payment systems face limitations in account setup, authorization, fees, and settlement efficiency. Stablecoins are particularly well-suited for such transactions and may become the primary transactional digital asset in AI agents’ business operations. Additionally, the AI computing power market could unlock new use cases for crypto assets. As AI enterprises’ demand for computing power continues to grow, computing power suppliers may in the future tokenize the rights associated with their computing capacity, allowing these tokens to be transferred, traded, or used as collateral, and AI agents can automatically purchase the required computing resources via related markets. BlackRock believes AI is poised to become a structural catalyst for digital asset adoption, while digital assets could also serve as critical infrastructure for the AI economy. Currently, this potential link remains underestimated.

The U.S. Treasury Department stated that it will repurchase at least $4 billion and up to $6 billion in long-term Treasury bonds on Thursday.

According to a Reuters report, a senior Iranian official said Tehran is reviewing the U.S. response to its proposal to end hostilities. During indirect talks with the U.S. on Tuesday, the two sides discussed reopening the Strait of Hormuz and lifting U.S. sanctions. Iran’s key priorities are a permanent end to hostilities and the lifting of the U.S. naval blockade. (Jinshi)

Polygon Foundation CEO Sandeep Nailwal announced in a post that the smart contract to burn 100 million POL tokens (approximately 1% of the total supply) is ready, and it can be triggered by any member of the community to execute. Going forward, the community will be able to carry out such burns on a quarterly basis. He also noted that POL has remained in a deflationary state so far this year.

According to an announcement from Silvia (formerly ProCap Financial, Nasdaq: SVIA), owned by crypto entrepreneur Anthony Pompliano, the firm will repurchase an additional 2.3% of its outstanding common shares at a price below net asset value (NAV) starting September 15. Since the repurchase program launched, the company has bought back roughly 14.5% of its total outstanding shares. As of the close of trading on September 22, Silvia held 82,881,223 outstanding shares, approximately 5,130 Bitcoin — a decrease of around 124 from the roughly 5,254 Bitcoin it held on September 15 — with a NAV of about $4.27 per share. The company said it will continue repurchasing shares until its stock price is no longer below NAV.