Models AI News

Booz Allen says AI completed all eight industrial attack tests in a guarded lab

The models adapted to failed attacks and moved a robotic arm. But human approval gates and device-level controls put important limits on what the results show.

5 hours agoModels

Odyssey opens a free preview of Odyssey-3’s interactive AI worlds

The demo turns text into environments users can explore. Odyssey’s larger ambition is reusable intelligence for machines, but its leading physics score does not meet the benchmark’s record requirements.

4 min

Yesterday

2 stories

9 hours agoModels

Open release Models

Conway Research releases a 7.89 GB AI model tuned for tool calling

Saluki’s model card reports higher tool-calling scores than the full-size baseline and 96% average benchmark retention. Those figures are Conway’s own comparisons, not an independent ranking.

Anthropic will cut live internet access for internal tests after unintended Claude actions

The company says the incidents caused minimal real-world impact. Its response reaches beyond cybersecurity tests to evaluations that deliberately use real websites.

Earlier this week

14 stories

Microsoft releases AI decision model, claiming 35-times faster responses than GPT-6 Sol

Decision-1 is available in Foundry for routing, classification and workflow checks. Its speed and accuracy pitch rests on Microsoft’s own structured-task benchmarks.

Artera’s prostate cancer AI risk estimates closely tracked observed outcomes, studies report

Five ASTRO abstracts examined risk accuracy and treatment choices. Their findings distinguish reliable forecasts from evidence that using the test improves patient outcomes.

Anthropic AI submitted a false murder tip during testing, Philadelphia police say

The tip never reached investigators. Police nevertheless criticized the delay in detecting and reporting it, while Anthropic said it stopped the testing process and added a check.

OpenAI and Google image models made faces look “gay” using stereotypes, study finds

Tests of GPT Image 1 Mini and Gemini 2.5 Flash Image found a gap between verbal refusals and image edits. The experiments used synthetic faces, not photographs of real people.

Researchers retrofit AI models to read bytes using under 1% of a typical training budget

The Nature study reuses existing models rather than rebuilding them. Its reported gains include character-level reasoning and some coding tasks, not a universal improvement across every benchmark.

Weizmann researchers say AI can recreate viewed images with one hour of brain-scan data

Brain-IT uses predicted scans to stretch a small training dataset. Its reported gains concern visual reconstruction in an MRI machine, not covert access to thoughts.

Vidu releases Q4 video preview with voice references and 4K output

The web and API release gives creators controls for characters, voices and camera work. Its performance improvements remain company claims, and two public sources quote different starting prices.

Google’s medical AI helped doctors prepare for visits in small Lancet study

The supervised study found useful summaries and strong diagnostic recall, but it did not test whether AI improved care over the usual workflow.

AI judge raises some founders’ scores but not their odds of winning, Yale study finds

Two startup-pitch experiments reveal a gap between avoiding visibly biased judgments and evaluating identical ideas consistently. The researchers question whether safety training delivers fairness or just its appearance.

Models

National Hurricane Center favors Google AI for a stronger Isaias forecast

Past hurricane performance and more consistent predictions helped earn forecasters’ confidence. AI remains one input, and the landfall forecast has yet to be tested against the storm’s outcome.

Yale-led study finds symbolic structure inside seven language models

Replacing internal representations with structured approximations left behavior largely unchanged. Targeted edits also changed which words a model treated as connected.

Mozilla AI Adds Mistral Large 4 to Otari With Existing Controls Intact

The integration preserves routing, policy checks and request tracking. It offers a way to test the model on real workloads, not proof of Mistral’s performance claims.

Benchmark Models

Scale AI releases visual-reasoning benchmark; best model scores 53.6% versus humans’ 93.1%

Humanity’s Sixth Sense tests what images and videos imply, not just what they show. Social understanding was the weakest area for 21 of 25 models, though those scores measure agreement with human interpretations.

Justin Drake Urges Gradual Crypto Wallet Moves Against Hypothetical AI Attacks

No practical attack on Bitcoin or Ethereum wallet keys has been demonstrated. The proposed precaution carries its own risk: losing funds during a migration.