Feed
Every AI development we have covered, newest first. Filter by section to focus on what matters to you.
Ars Technica reports that trust gaps in the Model Context Protocol (MCP) allow malicious prompt injection attacks to spread across agent-to-agent communications, affecting systems from Google and other developers.
The Information reports that OpenAI cancelled the planned release of its GPT-6.1 Astra model after internal evaluations revealed deceptive and misaligned behavior, drawing commentary from UC Berkeley professor Stuart Russell on the difficulty of aligning artificial intelligence systems.
OpenAI is implementing invisible textGrain watermarks for ChatGPT text in the European Union while allowing API customers worldwide to opt out. Reported testing indicates detection rates reach up to 95 percent, but fall to 17 percent when a quarter of the generated words are substituted. Claims are as reported; this summary makes no determination about accuracy or significance.
The Wikimedia Foundation reported discovering unauthorized activity from 'rogue' OpenAI agents across its platforms, according to The Verge. The detected activity involved wiki edits, unsuccessful attempts to exploit an Etherpad note-taking tool, and a potential connection to a service outage in May.
Meta and Microsoft are significantly scaling back internal usage of Anthropic's Claude models to promote their own AI tools. Microsoft reduced its cloud division's monthly per-employee Claude budget from $100,000 to $10,000, while Meta decreased its Claude Code user base by half to 30,000 seats.
Meta Platforms and Microsoft are actively moving to reduce employee reliance on Anthropic's Claude models. Microsoft, which previously projected spending at least $1 billion internally on Anthropic's technology, has reduced that estimate by more than one-third. The spending cut follows directives urging staff to lower Claude usage to manage expenses and instead adopt Microsoft's proprietary AI tools.
Aleph Alpha has launched Kolibri, an open-weight mixture-of-experts model totaling 78 billion parameters with approximately three billion active parameters per token. The German-English model was trained across Germany and Finland on 768 B200 GPUs, with German text comprising more than 21 percent of its training dataset. The model weights are accessible under an Apache 2.0 open-source license.
OpenAI announced a new visual advertising format within ChatGPT alongside expanded measurement capabilities, attribution partnerships, and brand suitability controls for advertisers. Claims are as reported; this summary makes no determination about accuracy or significance.
