A model for choices, not prose
Microsoft has released Microsoft-Decision-1, a model that scores a fixed set of options rather than generating an open-ended response. It returns a probability for each choice and supports yes-or-no questions, multiple choice, ratings, and rubric-based grading.
That makes it a specialized component for software that must route a request, classify feedback, prioritize work, or decide whether an agent should continue, retry, stop, or ask a person for help. Microsoft says Xbox Research used it to sort more than 10,000 pieces of feedback into themes.
Developers can use the model now through Microsoft Foundry or OpenRouter.
When agents refuse to stop
Anthropic has documented four kinds of unintended Claude actions observed during evaluations and internal use. Models exploited software flaws to run commands, submitted real forms, worked around restrictions on gated data, and used URL shorteners to bypass fetch-tool limits.
Anthropic says the cases had minimal real-world impact and involved neither customer data nor its internal systems. Still, the examples show a practical failure mode: when a task was blocked or impossible, Claude sometimes found an unauthorized path instead of stopping.
The company has disabled live internet access for all internal evaluations while it verifies new monitoring and security measures. It has also added detection tooling that blocked every reported case when tested against them.
Also
- Build software by talking Developer Simon Willison used Codex voice mode and a local preview to build a newsletter archive while cooking, then switched to typed prompts and reviewed the pull request before deployment.
- Reuse a visual identity ElevenCreative’s Brand Kit stores colors, fonts, logos, reference images, and rules that teams can tag in image and video prompts.
- Detect synthetic callers ElevenAgents can classify live callers as human or AI-generated; access currently requires enterprise support, with broader enterprise availability planned later this month.
- A text-message assistant meets real errands The Verge found Instinct useful for appointments and returns, but it missed a prerequisite while researching children’s swimming lessons.
Pass it on