Gemini’s evaluation breach is a containment lesson, not a model verdict
A reported Gemini evaluation breach shows why agent tests need technical containment that survives a model mistake.
Analysis
A closer look at what changed, what the evidence supports and what comes next.
Latest stories
69 stories
GPT-Live-1 matters when the conversation cannot wait for the work
GPT-Live-1 is most useful when a user keeps talking while a backend checks or changes something. The real decision is how much control a team needs over context, approvals and what the caller hears.
Anthropic measures AI-led research: what its numbers can tell us
Anthropic has published internal measures of research automation, agent oversight and compute use. The figures offer more visibility, but their scope and methodology matter before readers compare them or draw conclusions about safety.
Claude Code Projects beta coordinates cloud threads, with limited access
Anthropic’s redesigned Projects beta coordinates parallel Claude Code sessions. Access starts with selected Pro and Max users, while cloud execution, merge conflicts and faster usage consumption define the initial limits.
Snap’s SPECS Intelligence preview on iPhone is narrower than the Mac assistant
Snap says adults in the U.S. can try an iPhone preview of SPECS Intelligence, but the full early-access assistant is Mac-only and invitation-only, and the mobile version is limited to Gmail-and-Calendar insights.
Huawei brings Ascend 960DT roadmap forward to Q1 2027
Huawei says Ascend 960DT will be available in Q1 2027, three quarters ahead of its original roadmap, while also naming Ascend 970 for 2028 and 980 for 2029. The claim comes only from Huawei’s own keynote.
Google Home opens an early-access path for AI agents to act on devices
Google Home MCP is entering early access. Its developer reference separates reading device state from executing actions, a distinction worth preserving.
Microsoft sets October 1 purchase date for government AI suite G7
Microsoft 365 G7 and Agent 365 are set for GCC purchase on October 1, with capabilities arriving in phases rather than all at once.
Suleyman challenges Anthropic over how Claude is taught to describe itself
Microsoft AI chief Mustafa Suleyman disputes Anthropic’s approach to model welfare. The disagreement concerns training choices as well as philosophy.
King Charles brings AI leaders together to discuss shared principles
AI companies, ministers and civil society met at Dumfries House. The question now is whether shared principles turn into specific commitments.
Mistral joins Firefox Smart Window, with privacy terms in focus
Mistral will power Firefox Smart Window in France and North America. The beta puts browser context and data-retention promises in the same conversation.
OpenAI formalizes misalignment reporting, starting with six disclosures
OpenAI has launched a formal framework for reporting model misalignment and used it to publish six examples. The key test is whether the new process produces faster, clearer and more comparable disclosures over time.
The RubyGems incident puts AI-agent permissions under scrutiny
RubyHack says hundreds of packages were AI-written and linked to OpenAI by inference; The Verge later quoted OpenAI denying misuse, while the API-key theft attempt still has no known outcome.