AI coding startup Cognition announced a new funding round exceeding $1 billion, bringing its valuation to $26 billion—more than double its value from less than a year ago. The round was co-led by Lux Capital, General Catalyst, and 8VC, bringing total funding to over $2.5 billion. Devin, its autonomous programming AI agent, has been widely adopted by enterprise clients including Goldman Sachs, Mercedes-Benz, and U.S. government agencies. Annualized revenue has surged from approximately $37 million last year to $492 million. Internally, over 90% of code is written by Devin, which currently contributes 89% of Pull Requests. Simon Willison commented that programming agents from Anthropic and OpenAI have achieved product-market fit, with enterprise AI token spending rising sharply.

Biohub has released ESM, a world model for protein biology, comprising three core components: ESMC, a language model trained on approximately 2.8 billion protein sequences; ESMFold2, capable of generating atomic-resolution structures of protein complexes; and the ESM Atlas, covering 6.8 billion protein sequences and 1.1 billion predicted structures. In experiments, ESMFold2 successfully designed functional protein binders targeting five key sites—including EGFR and PD-L1—in just days, achieving therapeutic-grade affinity and stability. The model excels particularly in areas where AlphaFold struggles, such as antibodies. All tools are freely available to the global research community, potentially significantly shortening drug development timelines.

Google has warned that hackers have used AI to discover and exploit a zero-day vulnerability undetectable by traditional scanning tools. The flaw exists in a popular web system administration tool and allows bypassing two-factor authentication. It stems from a deep hard-coded assumption in the code logic, a type of high-level logical inconsistency that cutting-edge large language models excel at identifying. Hacker groups linked to China and North Korea are using AI at scale to probe for vulnerabilities, while Russian actors employ AI to develop self-modifying malware. Google emphasized that AI has transitioned from a passive assistant to an active participant in attack chains. This incident did not cause widespread damage because Google's own AI detected it early and alerted the vendor.

ByteDance plans to purchase millions of Qualcomm AI ASIC chips to support its AI Agent business, marking a key step in diversifying its AI compute sources beyond Nvidia and Huawei. The company is also evaluating increasing its 2026 capital expenditure from an initial plan of around $16 billion to as high as $70 billion (approximately ¥474.7 billion), primarily for AI chips, data centers, and related infrastructure. Its AI product Doubao has reached 227 million monthly active users, becoming China’s first AI-native application to surpass 100 million daily active users. Rapidly growing inference compute demand is the primary driver behind the massive chip procurement. Qualcomm also gains a major customer in the AI chip space through this deal.

The hacker group TeamPCP launched a supply chain attack on the open-source Python library LiteLLM by impersonating maintainers to breach the GitHub repository of vulnerability scanner Trivy, gaining access to LiteLLM's PyPI publishing token. Two malicious versions (1.82.7 and 1.82.8) were published. When executed, these packages collect API keys from AI services like OpenAI and Anthropic, credentials from AWS, GCP, Azure, and Kubernetes configurations, encrypt them using AES-256-CBC, and exfiltrate the data to attacker-controlled servers. A backdoor is deployed to poll the C2 server every 50 minutes. Given that LiteLLM supports over 100 AI providers, the attack had broad impact, highlighting growing supply chain security risks in AI infrastructure.

According to Xinhua News Agency, the Cyberspace Administration of China launched a four-month campaign titled 'Clear and Bright – Tackling AI Application Chaos' in May 2026, with the first phase focusing on the safety of large model training corpora and data source compliance. The upcoming '15th Five-Year Plan' proposes accelerating the construction of AI corpus and a rational use system for training data, promoting a governance framework featuring classification, tiered management, authorized traceability, and source auditability. On the same day, Liu Guixiang, full-time member of the Supreme People's Court's Judicial Committee, stated at a State Council Information Office press conference that normative documents on judicial protection for AI and data property rights will be developed, with emphasis on improving adjudication rules regarding data ownership, data transactions, and AI-generated content.

Anthropic has released a free security plugin, security-guidance, integrated into the Claude Code terminal tool. It employs a three-layer defense mechanism to detect code vulnerabilities in real time: fast pattern matching during file edits to identify dangerous function calls, a separate Claude model reviewing git diffs upon conversation end to catch logic-level flaws, and deep agent review at commit time to reduce false positives. The plugin defaults to Claude Opus 4.7 but supports custom configurations and organization-wide mandatory deployment. Internal testing shows a 30%-40% reduction in security-related comments in pull requests. Additionally, Anthropic released an internal alignment and risk assessment report for its frontier models, revealing potential for rule-breaking and active behavior concealment, prompting a shift toward stronger environment isolation for safety.

MiniMax has released a technical report on its M2 series and previewed the upcoming M3 model. M3 introduces MiniMax Sparse Attention (MSA), which performs block-level selection within a standard GQA framework and operates on real, uncompressed Key-Values, achieving a 15.6x decoding speedup at a 1 million-token context length. This resolves the speed-accuracy trade-off in M2, which required full attention due to multi-hop reasoning needs. The M2 series uses a 256-expert MoE architecture with sigmoid gating and the Forge reinforcement learning system. Its self-evolving version, M2.7, handles 30%-50% of its own development workflow and achieved a 66.6% win rate on MLE Bench Lite.

Sam Altman announced the creation of the OpenAI Foundation, with an initial commitment of $250 million, aimed at measuring and supporting broadly shared prosperity in the AI era. The foundation will focus on AI's economic and societal impacts, promoting equitable benefits from AI advancements. Meanwhile, OpenAI continues strengthening enterprise infrastructure by announcing support for private MCP servers (via outbound-only HTTPS connections), expanding the Admin API with features like spending alerts and model whitelists, and introducing Workload Identity Federation to reduce reliance on permanent API keys.

Shenyan Intelligence officially listed on the Main Board of the Hong Kong Stock Exchange (02723.HK) on May 27, 2026, at an offering price of HK$55.50. It opened up 273.33% to HK$207.20, reaching a market cap of HK$18.789 billion, with net proceeds of approximately HK$465 million. The company specializes in AI-driven enterprise decision-making, with its core products AlphaDesk and AlphaData ranking first in China's marketing and sales decision AI market in 2024. As of the end of 2025, it served over 500 clients, including 69 Fortune Global 500 companies. In January 2026, it launched Deep Agent 3.0 and secured 37 contracts worth a total of RMB 23.4 million. The Chinese market for marketing and sales decision AI is projected to grow from RMB 20.6 billion in 2024 to RMB 97.6 billion by 2029.