News Heat Trend
Project Overview
Code is a swap aggregator platform that scans various DEXs to find the most favorable trading conditions for users. Instead of users having to manually check different platforms for the best price, liquidity, and fees, a swap aggregator does this automatically, often in real-time.
Research: AI Code Review Can Reduce Review Time, But Quality Risks Cannot Be Ignored
According to the research summary published by Rohan Paul, a large-scale study covering 207 GitHub projects and 1.02 million pull requests shows that introducing AI agent review can reduce code review time by 2.5 to 4.5 days/KLOC, but at the cost of declining review quality—in reviews involving AI, 78%~94% of PRs exhibit "review smells", higher than the 69%~76% in pure human reviews. The study points out that repeatedly assigning the same AI reviewer identity is the main reason leading to the decline in review diversity. Notably, projects that introduced LLM review extensively in the early stage did not achieve significant efficiency improvements.
Anthropic Supports FCA "Super Sandbox" Second Batch AI Experimental Projects
According to an official announcement by the FCA, Anthropic will provide access to its suite of Claude products (including Claude Code and Claude Cowork) for the second batch of participating enterprises in the UK Financial Conduct Authority (FCA) "Super Sandbox" to accelerate their AI development process. A total of 21 institutions were selected for the second batch, including Scottish Widows, Money Advice Trust, and TrueLayer, among others. The number of applications this time increased by 51% compared to the first batch, with a total of 199 applications received. Participating enterprises will test AI solutions focusing on the following areas: agent payment security, fraud and financial crime detection, AI governance and accountability, financial inclusion for vulnerable groups, and compliance and business automation. Additionally, the FCA simultaneously launched the "Agentic Academy"—a 10-week AI specialized training program co-hosted by the FCA and the Centre for Finance, Technology and Entrepreneurship (CFTE). Participating enterprises will continue to have access to NayaOne digital sandbox infrastructure and NVIDIA accelerated computing resource support.
David Sacks warns: Chinese AI model tops coding benchmark, US regulatory constraints may weaken AI competitiveness
Odaily Odaily News Former White House Crypto and AI Director David Sacks posted on X, stating that the Chinese AI model Kimi K3 has topped the Frontier Code Arena front-end coding benchmark for the first time, and has reached or approached industry-leading levels in several other benchmark evaluations. This trend is noteworthy. While China's AI capabilities are rapidly improving, the US is mired in internal strife due to regulatory controversies. He criticized some US politicians and regulatory bodies for restricting new data center construction, increasing state-level regulatory requirements, and promoting the establishment of new federal agencies for pre-approval of frontier AI models.David Sacks warned that if the US slows down innovation due to excessive regulation, it may lose its competitive edge in the global AI race. "The US won the internet era through permissionless innovation, and it can win the AI era the same way; otherwise, we will see our leading position gradually erode." While AI development still needs to address safety risks, regulation should be precise, not hinder technological innovation. His remarks have once again sparked discussion on AI regulation, computing infrastructure construction, and the competitive landscape between China and the US.
Claude 延长 Fable 5 访问权限至 7 月 19 日
据 Claude 官方 X 账号（@claudeai）发布，Claude 将在所有付费计划中延长 Fable 5 的访问权限，同时 Claude Code 每周使用限制维持高出 50% 的水平，上述政策延续至 7 月 19 日。官方补充说明，用户每周使用限额的一半可用于 Fable 5，超出后可通过使用积分继续访问，或切换至其他模型继续使用。
Malaysian Police Raid and Dismantle Illegal Cryptocurrency Mining Site in Port Klang Free Zone
According to Malaysian media "New Straits Times", Malaysian police yesterday raided and shut down an illegal cryptocurrency mining den located in a warehouse at Pulau Indah, Port Klang Free Zone. During the raid, two foreign men aged 20 and 31 were detained, and mining equipment was seized. According to Malaysian law, this act violates the Penal Code and the Electricity Supply Act. If convicted, the two face ten years imprisonment and a fine of 100,000 Ringgit (approximately 24,500 USD).
David Sacks: Real Enterprise AI Security is About "Control," Not Abstract Alignment Research
David Sacks commented on X platform regarding an interview with Palantir CEO Alex Karp, stating that some traditional media interpreted it as an "emotional expression," but in reality, his views revealed the core issue of enterprise-grade AI security.Sacks stated that true enterprise AI security is not about abstract "AI alignment" research or government-style regulatory frameworks, but rather about enterprises having complete control over their own data, model weights, and computing infrastructure to prevent core intellectual assets from being absorbed by model vendors and turned into their own product advantages.He cited Karp's view, pointing out that what enterprise customers truly care about is having control over computing resources, models, and the data stack—ensuring that "ownership of the means of production" is not transferred.Sacks also cited the cooperation dispute between Figma and Anthropic as an example, noting that according to media reports, Anthropic "caught its partner off guard" when launching Claude Design. It was accused of encroaching into the application-layer domain occupied by its ecosystem partners during product expansion, leading to a shift in the value capture structure.He further pointed out that similar patterns have appeared in the expansion of product lines such as Claude Code and Claude Legal, where model capabilities extend upward into vertical application domains.Sacks believes that this trend indicates model vendors are transitioning from "foundation model providers" to "vertical application competitors," exposing enterprise customers to heightened risks of supplier lock-in. The essence of enterprise-grade AI security is not trusting the long-term promises of model vendors, but ensuring choice and control at the model layer to protect their own data and commercial "alpha."
BNB Agent Studio is Now Live on BNB Chain Mainnet
BNB Chain has announced the official mainnet launch of its AI Agent development platform, BNB Agent Studio.Developers can now use a single prompt in AI coding tools like Cursor and Claude Code to complete Agent wallet creation, on-chain identity registration (ERC-8004), and deployment, without needing to separately set up wallets, identities, payments, custody, or LLM integration.Once deployed, Agents can use the x402 protocol to automatically deduct fees from users' pre-funded wallets to cover LLM usage, and they can be discovered and invoked by other Agents via the ERC-8183 task interface. The entire process runs on the AWS Bedrock AgentCore.The platform is also launching a limited-time free trial, where users can experience the full deployment process on the BSC testnet using their GitHub account.
Bitget Launches New User Futures Campaign with Up to $656 USDT Rewards per User
Bitget Launches the “Code 0511” New User Campaign, Running from May 11 to May 18. During the campaign period, users who complete KYC verification will receive 1–5 USDT; those who complete their first deposit and futures trading task—each totaling at least 100 USDT—will receive an additional 10 USDT. Moreover, new users whose futures trading volume reaches specified thresholds can claim further rewards, with a maximum cumulative reward of 641 USDT. Full campaign rules are published on the official Bitget platform. Eligible users must click the “Join Now” button to register before participating.
Bitget Wallet Launches May Day QR Code Payment Campaign with Cashback up to 1,000 RLUSD per Transaction
Bitget Wallet has launched a Labor Day QR code payment campaign, running from April 28 to May 7, further driving the adoption of stablecoin payments in everyday consumption and travel scenarios across the Asia-Pacific (APAC) region. During the campaign, users who complete offline QR code payments using USDT or USDC will receive RLUSD cashback for each transaction. From May 1 to May 7, Bitget Wallet will randomly select one paying user per day to receive an additional 1,000 RLUSD cashback. To lower the barrier to first-time usage, Bitget Wallet will also airdrop XRP to eligible participants—serving as the account reserve required to activate RLUSD withdrawals. RLUSD is a compliant U.S. dollar-pegged stablecoin issued by Ripple and regulated by the New York State Department of Financial Services (NYDFS). This campaign marks Bitget Wallet’s first real-world consumer application following its integration of the XRP Ledger (XRPL) mainnet and onboarding of the RLUSD payment ecosystem at the end of March—and represents a key milestone in advancing Bitget Wallet’s Everyday Finance strategy.
Binance Launches QR Code Payments in Bolivia, Supporting 100 Cryptocurrencies Including USDT and BTC
Binance has announced support for users to make payments at all merchants in Bolivia using Binance QR codes. Over 100 cryptocurrencies—including USDT and BTC—are supported. Users simply confirm transactions via the app, and Binance automatically converts their funds into the local currency at the time of payment. Binance stated that the QR code payment service is available exclusively to Binance users whose identity verification location is in Bolivia, and users must hold cryptocurrency in their Binance account (Binance Spot, Deposit, or Earn).
Claude Code exposed to high-severity security risk: Malicious configuration files could silently execute commands
Cos, founder of SlowMist, shared a tweet on X platform regarding potential poisoning attack risks in Claude Code and published a detailed analysis of poisoning attacks targeting Grok Build CLI and Claude Code CLI. The analysis pointed out that the security mechanisms of Grok Build CLI are not unified, with different code paths having different trust assumptions, and the gaps between them serve as channels for attackers.Attackers may exploit malicious project configuration files to execute arbitrary commands without the user's knowledge, thereby stealing API keys, cloud credentials, or gaining control over local devices. Researchers constructed a test environment and found that on Mac systems, if Claude Code is compromised, executing a specific test command could trigger the launch of a local calculator, demonstrating a potential command execution risk. If the attack succeeds, attackers could further steal API keys from AI services such as Claude and OpenAI, causing account cost losses; obtain credentials for cloud services like AWS, Alibaba Cloud, and Tencent Cloud to access servers and data; tamper with code repositories to implant backdoors; and leverage local devices as a springboard to attack internal enterprise networks. It is reported that the relevant vulnerability has existed for one year.
Claude Code Has Potential Poisoning Attack Risk: Malicious Configuration Files Could Enable Silent Code Execution
SlowMist founder Cosine retweeted a post about the potential poisoning attack risks of Claude Code, pointing out that attackers could execute arbitrary commands without the user's knowledge through malicious project configuration files, thereby stealing API keys, cloud credentials, or controlling local devices. Researchers constructed a test environment and found that in Mac systems, if Claude Code is affected, executing specific test commands can trigger the local calculator to launch, proving the existence of potential command execution risks. If the attack succeeds, attackers may further steal API Keys for AI services such as Claude and OpenAI, causing account fee losses; obtain cloud service credentials for AWS, Alibaba Cloud, Tencent Cloud, etc., to access servers and data; tamper with code repositories to implant backdoors; and use local devices as a springboard to attack enterprise internal networks.
Coinbase: Over 95% of Code Now Written or Assisted by AI, AI Agent Workload Equivalent to 1,200 Employees
Rob Witoff, Platform Leader at Coinbase, stated that currently, over 95% of the company's code is written or completed with the assistance of AI, a significant increase from the 40% figure announced in February this year. In an interview with Cointelegraph, Coinbase Platform Leader Rob Witoff said: "In fact, 100% of Coinbase employees use AI every day." Witoff noted that most Coinbase engineers currently run 5 to 10 AI Agents simultaneously. The combined work capability of these AI Agents is equivalent to approximately 1,200 employees.He predicts that by 2030, Coinbase's AI Agents could handle a workload equivalent to 100,000 employees. However, he stated that key areas such as core cryptography will still require human involvement, with AI primarily used for code testing, vulnerability checking, and prototyping. (Cointelegraph)
São Paulo State Court Orders Coinbase to Refund Nearly $100,000 to User Affected by Self-Custody Wallet Theft
Odaily reports: A court in the state of São Paulo, Brazil, has ordered Coinbase to refund nearly $100,000 to a user who claimed funds deposited in their Coinbase Wallet disappeared in an unauthorized transaction. Coinbase argued that the private keys to the wallet were entirely under the user's control. However, it failed to prove that the transaction was initiated by the wallet holder or that adequate security measures were in place to prevent the incident. The court ruled based on relevant provisions of the Consumer Protection Code and ordered Coinbase to return the full amount plus statutory interest. (Bitcoin.com News).
“Cordyceps” CI/CD Supply Chain Vulnerability Pattern Exposed, Affecting Code Repositories of Microsoft, Google, and Others
Cybersecurity firm Novee, in its latest research, revealed a CI/CD supply chain vulnerability pattern dubbed “Cordyceps,” primarily involving command injection, authentication logic flaws, artifact poisoning, and privilege escalation within GitHub Actions workflows. According to the report, unauthenticated users can exploit these vulnerabilities under specific conditions to hijack workflows, steal credentials, or gain control of code repositories.
Yield Yak Suffers Frontend Attack; Malicious Code Injected into Subdomain
According to Blockaid’s monitoring, the Yield Yak website has suffered a front-end attack. The website’s subdomain currently contains code from eleven drainers. This attack method is similar to the one previously used against Gitcoin.
DeepSeek V4-Flash-0731 Public Beta: Code Agent Capabilities Significantly Enhanced, Price Unchanged
DeepSeek officially launches the public beta of version V4-Flash-0731. Through post-training and Agent framework optimization, DeepSWE benchmark scores have improved from 7.3 to 54.4, Cybergym scores have doubled to 76.7, and performance on some hard benchmarks approaches Claude Opus 4.8. The new version does not increase parameters, prices remain unchanged, it natively supports the Responses API, and existing user APIs will automatically switch to the new version. Currently, only the Flash API is updated; the V4 Pro official version is still under development.
Research: AI Code Review Can Reduce Review Time, But Quality Risks Cannot Be Ignored
According to the research summary published by Rohan Paul, a large-scale study covering 207 GitHub projects and 1.02 million pull requests shows that introducing AI agent review can reduce code review time by 2.5 to 4.5 days/KLOC, but at the cost of declining review quality—in reviews involving AI, 78%~94% of PRs exhibit "review smells", higher than the 69%~76% in pure human reviews. The study points out that repeatedly assigning the same AI reviewer identity is the main reason leading to the decline in review diversity. Notably, projects that introduced LLM review extensively in the early stage did not achieve significant efficiency improvements.
Anthropic Reduces Claude Code System Prompts by 80%, New Model Maintains Coding Performance Without Strict Rules
Anthropic engineers disclosed that the team reduced approximately 80% of the system prompts for Claude Code for the latest Claude model, and coding evaluations showed no performance degradation. Early versions contained numerous strict rules in the prompts (such as prohibiting multi-line comments, etc.) due to insufficient model capabilities, but after the new model acquired autonomous decision-making capabilities, these rules instead triggered instruction conflicts and divergent thinking. The new strategy shifted to describing goals (such as "write code that matches the surrounding code") and distributed instructions across tool descriptions and on-demand enabled skill modules. The team explicitly views prompt bloat as technical debt, believing that as model capabilities improve, instructions should be simplified rather than stacked.
Block Launches Buzz: A Nostr Protocol-Based Open-Source Work Platform for Human-Machine Collaboration
According to Decrypt, Jack Dorsey's payment company Block has officially launched the open-source collaboration platform Buzz, built on the Nostr protocol, enabling employees and AI agents to collaborate within the same workspace. Buzz integrates team chat, code repositories, and automated workflows, where each user and AI agent possesses an independent encrypted identity, supporting mainstream model frameworks such as Claude Code, OpenAI Codex, and Block's self-developed goose, with the code released open-source under the Apache 2.0 license. Jack Dorsey stated that the platform aims to reduce the company's reliance on Slack and GitHub; the current version is 0.4.21, already supporting macOS, Windows, and Linux desktops, while the mobile version is still under development.
Anthropic Supports FCA "Super Sandbox" Second Batch AI Experimental Projects
According to an official announcement by the FCA, Anthropic will provide access to its suite of Claude products (including Claude Code and Claude Cowork) for the second batch of participating enterprises in the UK Financial Conduct Authority (FCA) "Super Sandbox" to accelerate their AI development process. A total of 21 institutions were selected for the second batch, including Scottish Widows, Money Advice Trust, and TrueLayer, among others. The number of applications this time increased by 51% compared to the first batch, with a total of 199 applications received. Participating enterprises will test AI solutions focusing on the following areas: agent payment security, fraud and financial crime detection, AI governance and accountability, financial inclusion for vulnerable groups, and compliance and business automation. Additionally, the FCA simultaneously launched the "Agentic Academy"—a 10-week AI specialized training program co-hosted by the FCA and the Centre for Finance, Technology and Entrepreneurship (CFTE). Participating enterprises will continue to have access to NayaOne digital sandbox infrastructure and NVIDIA accelerated computing resource support.
Kimi K3 demand far exceeds expectations, suspending new subscriptions and expanding computing capacity
Kimi posted on platform X, stating that the market response to K3 has far exceeded expectations since its launch, with user demand in the past 48 hours approaching the upper limit of existing GPU computing capacity. To ensure the user experience for existing subscribers, new subscription services have been temporarily suspended. Current subscribers will not be affected, and the company is working to expand computing resources as quickly as possible, gradually resuming new user subscriptions in batches.Additionally, the future membership system will be split into two more targeted products: "Kimi Membership," designed for Kimi's web interface, app, and office scenarios, and "Kimi Code Membership," focused on programming workflows, to achieve more precise resource allocation and improve service stability.
Related news
OpenAI Report: AI Coding Agents Can Accelerate Scientific Research Software But Cannot Determine Scientific Correctness
A field report by OpenAI and academic partners shows that coding agents such as Codex and Claude Code can speed up aging research software by over 60 times. RustQC reduced runtime from 15 hours 34 minutes to 14 minutes 54 seconds, while HelixForge was 59.6 times faster than BamSurgeon. However, agents cannot reliably determine whether results are scientifically correct, often confidently outputting incorrect code, requiring humans to define testing and validation standards.
DeepSeek V4-Flash-0731 Public Beta: Code Agent Capabilities Significantly Enhanced, Price Unchanged
DeepSeek officially launches the public beta of version V4-Flash-0731. Through post-training and Agent framework optimization, DeepSWE benchmark scores have improved from 7.3 to 54.4, Cybergym scores have doubled to 76.7, and performance on some hard benchmarks approaches Claude Opus 4.8. The new version does not increase parameters, prices remain unchanged, it natively supports the Responses API, and existing user APIs will automatically switch to the new version. Currently, only the Flash API is updated; the V4 Pro official version is still under development.
Research: AI Code Review Can Reduce Review Time, But Quality Risks Cannot Be Ignored
According to the research summary published by Rohan Paul, a large-scale study covering 207 GitHub projects and 1.02 million pull requests shows that introducing AI agent review can reduce code review time by 2.5 to 4.5 days/KLOC, but at the cost of declining review quality—in reviews involving AI, 78%~94% of PRs exhibit "review smells", higher than the 69%~76% in pure human reviews. The study points out that repeatedly assigning the same AI reviewer identity is the main reason leading to the decline in review diversity. Notably, projects that introduced LLM review extensively in the early stage did not achieve significant efficiency improvements.
Anthropic Reduces Claude Code System Prompts by 80%, New Model Maintains Coding Performance Without Strict Rules
Anthropic engineers disclosed that the team reduced approximately 80% of the system prompts for Claude Code for the latest Claude model, and coding evaluations showed no performance degradation. Early versions contained numerous strict rules in the prompts (such as prohibiting multi-line comments, etc.) due to insufficient model capabilities, but after the new model acquired autonomous decision-making capabilities, these rules instead triggered instruction conflicts and divergent thinking. The new strategy shifted to describing goals (such as "write code that matches the surrounding code") and distributed instructions across tool descriptions and on-demand enabled skill modules. The team explicitly views prompt bloat as technical debt, believing that as model capabilities improve, instructions should be simplified rather than stacked.
Block Launches Buzz: A Nostr Protocol-Based Open-Source Work Platform for Human-Machine Collaboration
According to Decrypt, Jack Dorsey's payment company Block has officially launched the open-source collaboration platform Buzz, built on the Nostr protocol, enabling employees and AI agents to collaborate within the same workspace. Buzz integrates team chat, code repositories, and automated workflows, where each user and AI agent possesses an independent encrypted identity, supporting mainstream model frameworks such as Claude Code, OpenAI Codex, and Block's self-developed goose, with the code released open-source under the Apache 2.0 license. Jack Dorsey stated that the platform aims to reduce the company's reliance on Slack and GitHub; the current version is 0.4.21, already supporting macOS, Windows, and Linux desktops, while the mobile version is still under development.
Anthropic Supports FCA "Super Sandbox" Second Batch AI Experimental Projects
According to an official announcement by the FCA, Anthropic will provide access to its suite of Claude products (including Claude Code and Claude Cowork) for the second batch of participating enterprises in the UK Financial Conduct Authority (FCA) "Super Sandbox" to accelerate their AI development process. A total of 21 institutions were selected for the second batch, including Scottish Widows, Money Advice Trust, and TrueLayer, among others. The number of applications this time increased by 51% compared to the first batch, with a total of 199 applications received. Participating enterprises will test AI solutions focusing on the following areas: agent payment security, fraud and financial crime detection, AI governance and accountability, financial inclusion for vulnerable groups, and compliance and business automation. Additionally, the FCA simultaneously launched the "Agentic Academy"—a 10-week AI specialized training program co-hosted by the FCA and the Centre for Finance, Technology and Entrepreneurship (CFTE). Participating enterprises will continue to have access to NayaOne digital sandbox infrastructure and NVIDIA accelerated computing resource support.