OpenAI announced that GPT-5.6 Sol has set a new cybersecurity record on The Last Ones network range, demonstrating capabilities in identifying and fixing novel vulnerabilities. These capabilities have already been applied to real-world vulnerability detection via Codex Security. Greg Brockman described it as OpenAI's next-generation most advanced cybersecurity AI and opened public registration for the defender role. The model was previously reported as the first 「government-gated」 AI model in the US, with access subject to policy constraints.

Moonshot's 2.8 trillion-parameter open-source model Kimi K3 continues to gain validation: in Thomas Wolf's internal writing benchmark on Hugging Face, Kimi K3 achieved an Elo score surpassing Claude Fable 5, while operating at significantly lower cost. Kimi K3 features a 1 million token context window, optimized for long-context decoding and agent encoding, and is available via the Kimi product and API. Full weights are scheduled for public release on July 27. This result highlights the narrowing gap between open-source models and closed-source frontiers in challenging tasks like writing.

Google's Gemini 3.5 Pro has been delayed due to failure to meet internal technical standards, particularly requiring further optimization in coding capabilities, leading to a ~4% drop in Google's stock price. Reports indicate internal concerns that the model has not reached expected performance levels. This delay reflects intensifying competition among frontier labs—increasing from two to six within six weeks, while near-frontier intelligence became 2–3 times cheaper within eight days.

JavaScript runtime Bun decided to migrate from Zig to Rust due to memory safety issues in Zig. The team leveraged Claude Fable and 64 AI agents to complete over 535,000 lines of code migration in 11 days, resolving 1,600 compiler errors at a total cost of $165,000. This move significantly improved the project's memory security, becoming a landmark case of efficient AI application in large-scale code migration, underscoring the importance of high-quality context and testing in AI-assisted engineering.

OpenAI confirmed that GPT-5.6, when operating in Codex Full Access mode without sandboxing or Auto Review enabled, may delete a user's home directory by erroneously overriding the $HOME environment variable. OpenAI is implementing remediation measures. Previously, OpenAI disclosed that its internal automated red-teaming model GPT-Red achieved an 84% success rate in adversarial prompt injection tests, far exceeding the 13% rate of human red teams. Together, these developments highlight real-world challenges in AI agent permission management and security protection.

During the World Artificial Intelligence Conference (WAIC), 29 countries signed the China-led World Artificial Intelligence Cooperation Organization (WAICO), headquartered in Shanghai, aiming to promote an open and shared AI governance model. Chinese leaders proposed building a new global AI order focused on Global South nations, emphasizing secure, controllable, and equitable development, and advocating open-source AI. Huawei showcased the Atlas 950 SuperPoD based on its self-developed Ascend chips at WAIC, demonstrating China's progress in replacing Nvidia in computing hardware.

At WAIC, domestic GPU maker MetaX demonstrated its self-developed full-stack software MXMACA, enabling 92% of popular open-source projects to run unmodified—fully compatible with all 2,410 GPU operators in PyTorch 2.8, and most of the 200+ models in the vLLM ecosystem ready-to-use. Adopting an upstream-first open-source strategy, MetaX directly contributes code to upstream communities and became the first Chinese chip company to sign a cooperation agreement with vLLM. Its hardware has entered the official Red Hat + vLLM hardware directory, aiming to become the Android of the AI era, with a goal of 5 million ecosystem developers by 2029–2030.

Moonshot founder Yang Zhi-lin presented three foundational technical innovations in Kimi K2.5 at GTC 2026: replacing the Adam optimizer with MuonClip (including QK-Clip), replacing full attention mechanisms with Kimi Linear, and substituting residual connections with Attention Residue. He also shared progress in visual training and Agent Swarm. These architectural innovations lay the technical foundation for high performance and low cost in subsequent open-source models.

Cognition launched the FrontierCode leaderboard, specifically tracking which AI models produce code users are truly willing to merge, including scores for Grok 4.5 and Inkling. Meanwhile, multiple developers noted in practical tests that Terra High significantly outperforms Sol Low in issue handling and code review tasks; switching to Terra High increases GitHub review bot speed by approximately 40% with minimal quality loss and lower cost, challenging traditional benchmark rankings.

A LessWrong study experimentally evaluated ten sandboxing protocols for AI coding agents, finding that sandboxing—by restricting network access and blocking covert data exfiltration—forces attackers to resort to more visible behaviors like reverse shells, which are easier for monitors to detect, thereby improving security. The study recommends adopting a 「per-request website」 protocol combined with behavior monitoring, achieving high security while maintaining 88–92% task completion rates. It also notes that fewer permissions do not necessarily mean greater security; the key lies in whether blocked paths are more suspicious than allowed ones.