At its developer event, Perplexity unveiled the 'Personal Computer,' using a dedicated Mac mini as a persistent execution endpoint. This enables AI agents to access user files and applications 24/7 and perform tasks on their behalf—such as operating WhatsApp or Spotify—and return results upon completion. Simultaneously, Perplexity launched four new API categories: Search, Agent, Sandbox, and Embeddings, accelerating its shift from 'search' toward executable digital agents. The current setup remains geared toward developers and technical users, described as a cloud-based digital employee with access to personal data.

Reports indicate that Tencent's WeChat is exploring the development of an independent in-house AI model to build AI agent capabilities across its ecosystem. The model has completed foundational capability development and is scheduled for external deployment by 2026. The short-term strategy involves integration into the Mini Program ecosystem, opening access to developers for building various intelligent agents. Long-term goals include leveraging users' long-term behavioral data within WeChat to enhance information retrieval and productivity tools. With over 1.4 billion monthly active users, such deep integration also raises governance challenges around privacy protection and user trust.

OpenClaw released version v2026.3.11, fixing a potential cross-site WebSocket hijacking vulnerability under 'trusted-proxy' mode that could allow unauthorized sources to gain administrator privileges. The update strengthens permission controls for plugin runtime and sensitive interfaces. Meanwhile, China's CERT and the Ministry of Industry and Information Technology (MIIT) warned about weak default security configurations, recommending containerized isolation, disabling public network exposure, enforcing strict authentication and least-privilege principles, cautious use of skill/plugin markets, and vigilance against social engineering attacks. This coordinated action reflects that agent tools are entering a governance phase where usability and security are equally prioritized.

Google DeepMind and UC Berkeley introduced a reinforced 'generation' search framework in their paper 'Reinforced Generation of Combinatorial Structures: Ramsey Numbers' (arXiv:2603.09172), using LLM agents to automatically evolve search functions, replacing heuristic methods reliant on human expertise. The method introduces soft scoring signals like 'prospect graphs' to more effectively explore sparse extremal graph spaces, reducing the likelihood of getting trapped in local optima. Experiments show it reproduces nearly all known exact Ramsey numbers and achieves state-of-the-art performance in 28 distinct cases, demonstrating AI's capability for automated algorithm discovery in pure mathematics.

A Nature Medicine paper proposes a 'cognitive layer' architecture to enhance clinical reasoning in LLMs during psychotherapy conversations. The study conducted a randomized double-blind evaluation: 227 participants interacted with different therapeutic agents to generate dialogues, which were then reviewed by 22 clinical experts. Results showed that LLMs equipped with the cognitive layer consistently outperformed both standalone state-of-the-art LLMs and human therapists in key clinical competencies. Further analysis of 19,674 real conversation logs from 8,920 users revealed that higher activation of the cognitive layer correlated with greater symptom improvement and increased likelihood of long-term recovery (approximately 10 weeks). For safety and intellectual property reasons, the core code was not fully disclosed.

Nature Medicine introduces the Clinical Environment Simulator (CES), a 'digital hospital' framework for dynamically evaluating clinical LLMs, overcoming limitations of static datasets that fail to capture cascading effects and systemic constraints. CES consists of a 'hospital engine' that tracks bed, staff, and equipment status in real time, and a 'patient engine' that simulates disease progression and treatment responses under LLM interventions. Models must make decisions via real electronic health record interfaces, balancing individual treatment outcomes with system efficiency. The framework emphasizes evaluation across three capabilities: temporal reasoning, resource-aware decision-making, and operational resilience under concurrent emergencies and system failures, measuring both clinical and operational metrics.

Scale AI launched the FORTRESS benchmark to evaluate frontier large models' risk mitigation capabilities in national and public safety scenarios (NSPS). The benchmark includes over 1,010 expert-designed adversarial prompts (500 publicly available), covering three domains: CBRNE (chemical, biological, radiological, nuclear, and explosive), political violence and terrorism, and crime and financial illicit activities. It uses Average Risk Score (ARS) to measure propensity for generating harmful content and Over-Rejection Score (ORS) to assess false rejection of benign queries. The leaderboard shows Claude 3.5 Sonnet with an ARS of 12.96, while DeepSeek R1 scores 74.39. Evaluations are scored automatically by a multi-model adjudication system, emphasizing scalability and reproducibility.

AI video generation company AIsphere secured $300 million in Series C funding led by CDH Investments, marking a new record for single-round financing in China's AI video generation sector. The company reported over $40 million in annual recurring revenue (ARR) for 2025, with its overseas-focused PixVerse app and related products surpassing 100 million cumulative users. Funds will support ongoing R&D and global consumer market expansion. The report notes AIsphere launched PixVerse R1, a real-time world model enabling real-time video generation and 'infinite visual continuation,' in January 2026. Competitively, the company faces pressure from OpenAI's Sora and multiple domestic video modeling products.

Atlassian announced a global layoff of 1,600 employees, approximately 10% of its workforce, including around 480 positions in Australia. CEO Mike Cannon-Brookes stated that the cost savings will be redirected toward 'self-funded investments in artificial intelligence and enterprise sales.' He acknowledged that AI has reshaped required skill sets and job roles, with AI-related skills being prioritized in retention decisions. Anonymous employee sources cited prior over-hiring as a contributing factor. Market-wise, the company's stock price has dropped from $221 to $75.45 over the past year, reflecting transformation pressures on SaaS companies amid AI disruption and capital market repricing.

According to Japanese media reports, China's National People's Congress has approved the '15th Five-Year Plan,' an economic mid-term target extending to 2030. The plan emphasizes enhancing manufacturing competitiveness under external supply chain pressures and advancing the industrialization of scientific achievements in AI and semiconductors, aiming to grow the AI-related industry scale to over 220 trillion yen by 2030. Reports note more cautious wording in external communications, reflecting the need to balance technological self-reliance with international relations and supply chain risks. This target provides clearer policy expectations and industrial momentum for domestic computing power, chips, software, and application deployment.