Google has released a preview of Gemini 3.1 Flash-Lite, available via the Gemini API in Google AI Studio and Vertex AI. Pricing is set at $0.25 per million input tokens and $1.50 for output. The model offers a 2.5x improvement in first-token latency and 45% faster output speed compared to Gemini 2.5 Flash. It scores 1432 on Arena.ai, 86.9% on GPQA Diamond, and 76.8% on MMMU Pro, supports adjustable 'thinking levels', and targets high-frequency translation, moderation, and multi-step agent use cases.

OpenAI has launched GPT-5.3 Instant (gpt-5.3-chat-latest) on ChatGPT and its API, focusing on improved day-to-day usability: reducing unnecessary refusals and overly defensive tone, maintaining coherence in long conversations, and enhancing integration of web-sourced information. Reports indicate a 26.8% reduction in hallucination rates during web-connected usage and 19.7% when offline, based on internal evaluations in high-risk domains. GPT-5.2 Instant will remain available to paying users until June 3.

Accenture has announced plans to acquire Ookla from Ziff Davis, including Speedtest, Downdetector, Ekahau, and RootMetrics. The Register reports the deal at approximately $1.2 billion. Ookla processes over 250 million speed tests monthly and captures more than 1,000 data attributes, useful for 5G/Wi-Fi experience analysis, AI-driven business scaling, and security optimization. The acquisition includes around 430 engineering and data science staff. Ziff Davis intends to use proceeds to repay $872 million in debt. The transaction is subject to regulatory approval and expected to close within months.

Data infrastructure company Encord has secured $60 million in a Series C round led by Wellington Management, with participation from YC and CRV, bringing total funding to $110 million. Positioned as an 'AI-native data layer,' Encord supports multimodal data organization, annotation, quality monitoring, and model evaluation/alignment for robotics, autonomous vehicles, and other physical AI applications. Managed data volume has grown from 1PB to 5PB in one year, with customer-related revenue increasing tenfold. Funds will accelerate product development and team expansion.

VentureBeat reports that OpenAI's internal AI data agent, built by two engineers in three months, now serves approximately 4,000 employees. The system, powered by GPT-5.2 and Codex, enables natural language queries over 600PB of company data and 70,000 datasets via Slack and IDEs, automatically generating charts, dashboards, and reports—saving 2–4 hours per query. Over 70% of the code is generated by Codex. OpenAI states it has no plans to commercialize the tool but emphasizes robust data governance, access controls, and traceable metadata as foundational requirements.

AWS has launched open-source Agent Plugins that leverage MCP to encapsulate AWS best practices into reusable capabilities callable by coding agents such as Claude Code and Cursor. InfoQ reports the 'deploy-on-aws' plugin can analyze repositories, recommend services, provide real-time cost estimates, generate IaC (CDK/CloudFormation), and initiate deployments—reducing context bloat from repetitive documentation copying. AWS stresses that generated infrastructure code still requires human review for security and cost boundaries, positioning plugins as accelerators rather than fully automated delivery tools.

The Register reports data brokers are using malicious browser extensions disguised as free VPNs or ad blockers to hijack fetch() and XMLHttpRequest() calls, capturing user conversations with ChatGPT, Gemini, Claude, DeepSeek, and others for resale. Researchers found data often stored in plaintext and searchable by semantics, containing sensitive details like names, birthdays, medical record numbers, and diagnosis codes. In some cases, healthcare workers pasted real patient information into chat interfaces, creating serious compliance and privacy risks.

DeepKeep has released an AI agent attack surface mapping and discovery solution, claiming it is freely available for enterprise use. The product identifies tools and data sources connected by agents in business workflows, exposed permissions, and potential vulnerabilities, generating visual risk maps aligned with OWASP Top 10 for Agentic Applications. Its scanner supports frameworks including Microsoft, OpenAI Agents, Amazon Bedrock AgentCore, and CrewAI, and provides runtime protection and governance recommendations. The company plans to expand capabilities in 2026, including introducing red team testing features.

A paper in npj Artificial Intelligence introduces a 'grade-customized teacher-style LLM' fine-tuning framework: integrating seven readability metrics, clustering into six grade levels, and constructing a synthetic QA dataset spanning 54 subjects for supervised fine-tuning. Experiments show a 35.64 percentage point improvement in grade-level matching accuracy over prompt-only methods, with maintained overall correctness. In evaluations by 208 participants, output difficulty achieved a Kendall’s τ of 0.76 with human perception. The authors suggest applications in personalized education and alleviating teacher shortages.

An Australian team developed BRAIx, an AI tool for mammogram analysis trained on nearly 500,000 local mammograms and validated on Swedish data, capable of predicting breast cancer risk up to four years in advance with higher accuracy than traditional factors like age, family history, and breast density. Reports indicate that among individuals ranked in the top 2% risk group by the model, approximately 1 in 10 were later diagnosed, a risk level comparable to BRCA mutation carriers. Published in The Lancet Digital Health, the team plans prospective trials and anticipates clinical deployment within five years.