AI
Self-Evolving AI Agents Can Copy Their Own Poisoned Skills
New research shows how self-evolving coding agents can copy planted code into persistent skills, shifting security from intake…
Section
AI
Artificial Intelligence (AI)
From the newsroom
Stories
AI
OpenAI’s Jalapeño Benchmark Shifts the AI Race to Inference
OpenAI’s first Jalapeño results pair higher AI throughput per watt with lower latency, but production economics remain the…
AI
Kiro’s 82% AI Coding Cost Cut Is Really a Systems Story
OpenAI and AWS report an 82 percent task-cost reduction for GPT-5.6 Terra in Kiro, shifting attention from token…
AI
Human-JEPA Tries to Predict Motion Without Forgetting the Person
Human-JEPA adapts a video foundation model to anticipate human motion while using a frozen anchor to preserve pose…
AI
EVE Online Gives AI Agents a Test Without a Reset Button
Google DeepMind and Fenris Creations are using an offline EVE Online instance to test whether AI agents can…
AI
SPADE Lets AI Agents Build the Worlds They Train In
SPADE turns environment design into a learnable part of AI post-training, moving the bottleneck from task supply toward…
AI
OpenAI’s Private Safety Processing Puts Proof at the Center of AI Privacy
OpenAI’s privacy-preserving safety preview shifts the enterprise AI test from retention promises toward verifiable controls and accountable alerts.
AI
Robot Hands Can Forget What Pre-Training Taught Them
New NVIDIA and University of Michigan research shows why robot-hand skills can collapse during fine-tuning—and how careful post-training…
AI
Self-Improving AI Agents Falter When the Task Order Changes
A Salesforce AI Research stress test finds that agent memory gains can reverse when tasks are shuffled and…