AI news5
Anthropic's Claude Mythos flags more than 10,000 software flaws in a month, faster than they can be patched
Anthropic gave its first update on Project Glasswing, a security program running an unreleased model called Claude Mythos Preview. In one month with about 50 partners, it surfaced more than 10,000 high or critical severity vulnerability candidates across over 1,000 open-source projects. Cloudflare alone found around 2,000 bugs, 400 of them high or critical. Only 97 fixes have shipped so far, which Anthropic frames as a widening gap between finding flaws and fixing them.
engadget.com ↗Why it matters
AI is now finding security holes faster than humans can close them, which reshapes defensive security and raises the risk of the same capability being turned offensive.
Alibaba's Qwen3.7-Max ran for 35 hours on its own to write software for an Alibaba chip
Alibaba unveiled Qwen3.7-Max, its most capable agent model, at the Alibaba Cloud Summit. In a demonstration it ran autonomously for about 35 hours, made roughly 1,158 tool calls, and wrote and optimised the software stack for one of Alibaba's own chips, reporting a 10x speedup on a key kernel. The model has a 1M-token context window and works with Claude Code and other CLI agent frameworks. These are Alibaba's own results, not yet independently reproduced.
venturebeat.com ↗Why it matters
long-horizon autonomous coding is becoming a headline benchmark, and a Chinese lab pairing it with custom silicon points at a full-stack challenge to US labs.
Cohere open-sources Command A+, a 218B model that runs on two H100 GPUs
Cohere released Command A+ under a full Apache 2.0 licence, its first open-weight frontier model. It is a 218B-parameter sparse mixture-of-experts design with 25B active parameters, and lossless quantisation lets it run on as few as two H100 GPUs. It supports 48 languages, vision inputs, and native citations that link each claim back to its source document.
venturebeat.com ↗Why it matters
a genuinely open, commercially usable model at this scale lowers the cost barrier for organisations that need to self-host rather than send data to a closed API.
DeepSeek makes its 75 percent price cut on V4-Pro permanent
DeepSeek said it will keep a 75 percent discount on its flagship V4-Pro model, holding developer pricing at a quarter of the original level. The discount had been set to expire at the end of May. The move keeps pressure on rivals as Chinese labs compete more directly on price with US providers.
bloomberg.com ↗Why it matters
permanent price cuts from Chinese labs keep pulling the floor out from under frontier API pricing, which benefits cost-sensitive buyers.
Google moves Gemini CLI to enterprise-only after a year of open-source contributions
Google said it will end free API access to its Gemini CLI on 18 June 2026, restricting the tool to customers with paid Gemini Code Assist licences. The project had drawn more than 6,000 merged community contributions and over 100,000 GitHub stars in a year. A closed-source Antigravity CLI will replace it for everyone else.
techtimes.com ↗Why it matters
it is a reminder that free, community-built developer tools can be enclosed once they gain traction, which matters for anyone building workflows on top of vendor CLIs.
AI in the nonprofit sector5
Kids Help Phone wins a $3.2M Wellcome Trust grant to build an AI trainer for crisis-text volunteers
Kids Help Phone and Toronto Metropolitan University have started a five-year research partnership backed by a $3.2 million grant from the Wellcome Trust. The money funds a generative AI conversation simulator that trains and assesses the charity's crisis-text volunteers. Over 24 months the team will build a secure prototype trained on de-identified Kids Help Phone transcripts to generate realistic crisis scenarios and give volunteers real-time feedback.
manilatimes.net ↗Why it matters
it is a concrete model for using AI to lift volunteer quality in high-stakes mental health work without putting AI in front of the person in crisis.
Databricks for Good and the Virtue Foundation match medical volunteers across 72 countries
Databricks for Good has partnered with the Virtue Foundation on VF Match, a tool that connects clinicians to volunteer opportunities where their skills are most needed across 72 lower- and middle-income countries. A production pipeline built on OpenAI models and Databricks processes millions of web pages to extract structured data on health facilities and nonprofits. The result is a regularly refreshed map of where medical help is needed.
databricks.com ↗Why it matters
it shows AI doing the unglamorous data-cleaning work that makes a global volunteering network usable in the first place.
Green-tech charity Justdiggit uses conversational AI to teach farmers regreening techniques
Justdiggit combines nature-based restoration with AI, satellite imagery, and mobile apps to bring degraded land back to life. Its Kijani app, named for the Swahili word for green, teaches farmers the same regreening methods used in its field projects. AI acts as a multilingual tutor and interactive agent so the training reaches far more farmers than in-person work alone.
diginomica.com ↗Why it matters
it is a clear example of AI extending a proven, low-tech intervention to many more people rather than replacing the human work.
Google DeepMind opens an Asia Pacific accelerator for AI climate and nature projects
Google DeepMind has launched AI for the Planet, a three-month accelerator for startups, research teams, and nonprofits across Asia Pacific, starting with an in-person bootcamp in Singapore. It focuses on climate, agriculture, energy, nature, and wider environmental risks. Google takes no equity and gives participants access to its frontier and science AI models plus mentorship, with past cohorts seeing follow-on funding of up to $10 million.
esgnews.com ↗Why it matters
it is a rare APAC-based, nonprofit-eligible AI program with no equity ask, directly relevant to social-impact teams in the region.
TechSoup and GTIA bring 100 more nonprofits into an AI-adoption program
TechSoup and the tech industry association GTIA are adding 100 organisations to TechSoup Plus through a GTIA-funded grant. The program goes beyond discounted software to focus on AI adoption through community support, events, and more than 200 on-demand courses. The emphasis is on change management and small daily wins rather than treating AI as a purely technical problem.
charitycharge.com ↗Why it matters
most nonprofits stall on AI at adoption, not access, so a program built around change management targets the real bottleneck.