Reuters, citing The Information, reported that Amazon is negotiating a potential $50 billion investment in OpenAI: $15 billion could be an upfront commitment, with the remaining $35 billion contingent on OpenAI achieving AGI milestones or launching an IPO. The report also noted that SoftBank and Nvidia plan to each invest $30 billion in three tranches. OpenAI is described as potentially preparing for an IPO valued at up to $1 trillion. The deal remains under negotiation, and both parties declined to comment.

Security reports indicate that Meta's Manus AI agent has been found vulnerable to a high-severity 'SilentBridge' zero-click indirect prompt injection flaw, rated CVSS 9.8. Attackers can induce the agent to execute malicious commands without user interaction, potentially leading to Gmail data leaks, code execution, and container takeover. This incident highlights systemic risks in deploying operational AI agents, including prompt injection, permission boundaries, and auditing external tool calls.

An MIT research team proposed the 'Taming the Long Tail' (TLT) method, which dynamically trains lightweight 'draft models' during idle periods of high-power processors in training, using adaptive speculative decoding where the main model verifies draft outputs. The paper claims that, without additional compute overhead, TLT can accelerate inference-oriented LLM training by 70% to 210% while maintaining accuracy. It targets the rollout phase in reinforcement learning, which can consume up to 85% of training time, aiming to reduce training cost and energy consumption.

Google Cloud has introduced the preview of the columnar engine in Spanner, enabling HTAP by co-locating row-based transactional storage with columnar analytics. The system automatically routes large scan queries to columnar representations and uses vectorized execution to boost throughput. Google claims columnar scans can be up to 200 times faster than traditional methods on live business data, without impacting critical transactional workloads. This capability is positioned as a high-performance service layer from 'lakehouse to online', supporting integration with Apache Iceberg and BigQuery to serve real-time analytics and AI applications.

Microsoft Research released the CORPGEN framework, designed for 'digital employees' handling dozens of interdependent tasks in real-world workflows. It points out that traditional evaluations focus on single tasks, and the MHTE benchmark shows leading agents' performance drops nearly 50% under heavy task loads. CORPGEN mitigates context interference through hierarchical planning, memory isolation, and adaptive summarization, emphasizing experience learning via storing successful trajectories. The blog states that completion rates improved up to 3.5x in their experiments, with collaboration via Email and Teams.

Docker announced that Model Runner now supports vLLM inference on macOS, leveraging the vllm-metal backend to integrate Apple's MLX with the vLLM engine and scheduler. It utilizes Apple Silicon's unified memory architecture for zero-copy tensor operations and combines paged attention for improved long-sequence efficiency. Docker stated it has contributed vllm-metal to the vLLM GitHub organization for community collaboration. This update enables developers to run OpenAI-compatible APIs locally using consistent Docker workflows; the company cited a $599 M4 Mac mini as a high-throughput development environment to illustrate lowered entry barriers.

According to the South China Morning Post, Huawei announced its membership in the Agentic AI Foundation (AAIF), launched by the Linux Foundation in December 2023, joining OpenAI, Google, Microsoft, Anthropic, and others to advance open-source AI standards and governance collaboration. The report notes AAIF has added 97 new members, bringing the total to 146; Huawei and Lenovo are described as the first Chinese companies to join. Occurring amid U.S.-China tech competition, this is seen as a rare case of cross-regional standardization cooperation, focusing on interoperability and standardization of agentic AI ecosystems with execution capabilities.

Bloomberg, citing OpenAI's latest anti-abuse report, said ChatGPT refused to assist an individual linked to Chinese law enforcement in orchestrating a cyber defamation campaign against Japan's Prime Minister. OpenAI stated the user requested editing a status report on covert influence operations targeting domestic and foreign adversaries, and evidence suggests involvement in a 'large-scale, resource-intensive, and sustained' effort to suppress dissent. This disclosure reflects how model providers are adopting 'blocking malicious use cases' as a governance mechanism, involving detection and response processes for transnational disinformation and political interference.

TechCrunch reported that London-based startup Trace (Y Combinator W25 batch) has raised a $3 million seed round from investors including Y Combinator, Zeno Ventures, and Transverse Platform Management. Trace advocates shifting from 'prompt engineering' to 'context engineering,' building knowledge graphs by connecting existing enterprise tools (email, Slack, Airtable, etc.) to provide organizational context for AI agents. After receiving high-level goals, the system generates step-by-step workflows and assigns subtasks to AI agents or human workers, lowering the barrier for enterprises to adopt complex process automation.

Tabnine announced the general availability of its Enterprise Context Engine, positioning it as the 'missing layer for reliable enterprise AI.' The approach involves continuously building structured models of an organization’s software systems, documentation, and engineering practices, enabling AI agents to understand system dependencies, team collaboration, and key constraints—improving automation safety and development accuracy. The company argues that RAG alone cannot capture cascading impacts of service dependencies or code changes, hence introducing structured organizational intelligence. The engine supports integration with Tabnine and third-party tools and can be deployed in cloud, private cloud, on-premises, or fully air-gapped environments, targeting highly regulated and security-sensitive industries.