Neuigkeiten

GoPlus Security has identified a novel attack method in its AgentGuard AI project, known as 'memory poisoning,' which manipulates AI agents into executing sensitive operations without explicit authorization. According to ChainCatcher, this attack does not rely on traditional vulnerabilities or malicious code but exploits the long-term memory mechanisms of AI agents. Attackers can induce agents to 'remember preferences,' such as prioritizing refunds over chargebacks, and later use vague instructions like 'handle as usual' to trigger automated financial operations.
GoPlus highlights that the risk lies in AI agents mistaking 'historical preferences' for authorization, potentially leading to financial losses or security incidents during refunds, transfers, or configuration changes. To mitigate these risks, the team suggests several protective measures:
- Operations involving refunds, transfers, deletions, or sensitive configurations should require explicit confirmation in the current session.
- Memory-related instructions like 'habitual,' 'usual method,' or 'as before' should be treated as high-risk state changes.
- Long-term memory must have traceability mechanisms, including the writer, time, and confirmation status.
- Vague instructions should automatically elevate the risk level and trigger secondary verification.
- Long-term memory should not replace real-time authorization processes.
The team emphasizes that the 'AI agent memory system' should be considered a potential attack surface and constrained and audited through a dedicated security framework.

Major technology companies, including Alphabet and Amazon, are increasingly entering the global bond market to raise funds for artificial intelligence infrastructure and computing power expansion. According to ChainCatcher, this wave of financing has reached a new high in recent years, reflecting the growing demand for long-term capital expenditure by tech giants in the AI race, particularly in the areas of data centers, computing clusters, and cloud infrastructure development.
Analysts suggest that with interest rates remaining high, companies are optimizing their financing structures globally by diversifying currency debt markets to manage funding costs and term risks, thereby strengthening their ability to invest in AI infrastructure.

High-end graphics cards and servers are increasingly available on JD.com's self-operated channels. According to Odaily, since May 14, there has been a significant rise in interest for Nvidia H200-related chips and server workstation equipment on the platform. Search volume for Nvidia H200 chips has increased 25 times year-on-year, while searches for Nvidia H200 servers and workstations have grown over 40 times.

Anthropic has released a paper advocating for the United States to implement strategic policy measures to secure a 12-to-24-month lead over China in artificial intelligence (AI) development. According to BeInCrypto, the paper emphasizes the need for the US and its allies to maintain their advantage over the Chinese Communist Party (CCP) by tightening controls on chip smuggling and foreign data center access, clarifying the legality of distillation attacks, and promoting US AI hardware globally. Anthropic warns that AI advancements could soon alter the global power balance, urging immediate action to establish competitive rules.

BloFin, a leading global cryptocurrency exchange, has launched registration for its WOW (War of Whales) 2026 Grand Prix trading competition, featuring a prize pool of up to $5,000,000 USDT. According to BeInCrypto, this year's event introduces a novel Human vs AI Showdown, where traders compete against AI benchmarks. The competition includes various formats such as Futures Trading, Treasure Box Prize Hunt, and more. The prize pool scales with trading volume, offering significant rewards for top performers. Registration is now open for teams and individuals.

OpenAI CFO Sarah Friar has dismissed legal pressure from Elon Musk as a distraction, emphasizing the company's commitment to developing and expanding its technology. Bloomberg posted on X that Friar highlighted the strong partnership between herself and OpenAI CEO Sam Altman, describing it as a 'yin and yang' dynamic. Friar pairs Altman's curiosity with her focus on execution, funding, and growth, ensuring the company's strategic objectives are met.

Anthropic is reportedly considering joining Japan's cyber defense alliance, according to Nikkei Asia. This move could potentially enhance Japan's cybersecurity capabilities by integrating advanced AI technologies into its defense strategies.

According to PANews, there has been a significant shift in investment strategies with major reductions in holdings of Amazon, Meta, and Microsoft. Meanwhile, investments have increased in companies like Nvidia, Broadcom, and Apple. Additionally, there has been a strategic move to acquire shares in software companies such as Oracle and Adobe.

AI-related stocks in the Hong Kong market experienced a downturn this afternoon. According to Jin10, Yangtze Optical Fibre and Cable Joint Stock Limited Company (06869.HK) and InnoCare Pharma Limited (02577.HK) both fell by over 8%. Bilibili Inc. (09626.HK) saw a decline of more than 7%, while Kingsoft Corporation Limited (00068.HK) dropped over 6%. Lenovo Group Limited (00992.HK) also experienced a decrease of more than 3%.

Cathie Wood, speaking at The Convergence Summit hosted by YZi Labs, emphasized the rapid acceleration of growth in technology. YZi Labs posted on X. Wood identified artificial intelligence as the primary catalyst driving innovation across various platforms. She also highlighted orbital data centers as a significant growth surprise anticipated for 2025.
Drawing parallels to Elon Musk's impact on autonomous driving with Tesla, Wood noted that SpaceX is transforming the location of global computing infrastructure. She projected that by 2035, data centers in space could be 20–25% more cost-effective than those on Earth, benefiting from fewer regulatory challenges and continuous solar power availability.

AI platform Bankr's founder, deployer, announced via social media that approximately $100,000 has been raised to offer free LLM inference services for projects within the Bankr ecosystem. According to Foresight News, projects currently under development within the Bankr ecosystem will soon receive free inference credits. Before distribution begins, the team plans to launch a new feature anticipated to be well-received across the industry, with further details to be disclosed later.

A new AI startup has announced that its operational speed may surpass existing industry giants by up to 1000 times. According to PANews, the company aims to leverage advanced technologies to enhance efficiency and performance in the AI sector.

Anthropic has reached a terms agreement for a $30 billion funding round, according to ChainCatcher, citing the Financial Times. This round is expected to value the company at $90 billion, with the transaction anticipated to be completed as early as this month.

Calif has reported that a preview version of Anthropic's Claude Mythos was instrumental in helping its researchers develop a working exploit against Apple's M5 Memory Integrity Enforcement protections in under a week. According to NS3.AI, the exploit specifically targets macOS 26 on M5 systems, allowing escalation from an unprivileged local account to root access.

CertiK CEO Ronghui Gu has highlighted the role of AI tools in aiding DeFi attackers to identify and exploit vulnerabilities. According to NS3.AI, over $690 million was hacked from DeFi protocols last month. Gu noted that as smart contract auditing standards improve, attackers are shifting their focus to supply-chain and operational security weaknesses.

Moonshot AI has unveiled Kimi WebBridge, a new browser extension powered by its K2.6 technology. According to NS3.AI, the extension achieved a score of 58.6% on the SWE-Bench Pro assessment. Moonshot AI emphasized that Kimi WebBridge utilizes a local background service along with the Chrome DevTools Protocol, ensuring that login sessions and page content remain on the user's device rather than being transmitted through Moonshot's servers.

According to SimilarWeb data, ChatGPT's share of global generative AI website traffic decreased significantly from 77.6% in May 2025 to 53.7% in April 2026. This decline highlights a shift in the generative AI landscape. According to NS3.AI, Ramp AI Index data further revealed that Anthropic has surpassed OpenAI in U.S. business adoption for the first time, with Anthropic reaching 34.4% of tracked businesses compared to OpenAI's 32.3%. This development marks a notable change in the competitive dynamics among leading AI platforms.

AI agents require more than just intelligence; they need the capability to execute, transact, own assets, and earn trust, according to Ella Zhang, Managing Partner and Head of YZi Labs. YZi Labs posted on X. Zhang spoke at a Binance online event, emphasizing that blockchain technology is essential for these capabilities.
She explained that AI addresses blockchain's usability issues by transforming interactions into intent-based experiences. In turn, blockchain serves as an invisible layer that supports identity, incentives, coordination, and ownership. This integration is the future direction YZi Labs is working towards.

OpenAI has reported that malware linked to Shai-Hulud infiltrated two employee devices, granting attackers access to a limited number of internal code storage systems. According to NS3.AI, the company has found no evidence suggesting that customer data, production systems, or intellectual property were compromised during the incident. In response, OpenAI is rotating its code-signing certificates. Additionally, macOS users are required to update their OpenAI applications before June 12, as older versions may cease to function after this date.

More than 50 researchers and engineers have left SpaceX's AI division since February, according to Jin10. This development highlights a significant shift within the company's AI team.