Microsoft researchers report that Lazarus used seven edited neurons to trigger a command in an AI agent, while leaving ordinary tests largely unchanged.
Browsing Category
AI & Tooling
1323 posts
Google And CSIRO Explore Whether AI Can Help Tasmania’s Giant Kelp
Google says AI helped analyze 600 wild giant kelp specimens in a breeding project. The peer-reviewed genome work is published; survival at sea remains untested.
What Microsoft Discovered About 7-Neuron AI Sleeper Agents
Microsoft researchers say Lazarus reprogrammed seven neurons in one AI model to trigger a hidden command while routine benchmark performance remained close to the original.
The Data Science Skills Consultants Showed With ChatGPT Disappeared Without It
A randomized BCG trial found ChatGPT raised consultants’ data science task scores, but a follow-up test found no measurable knowledge gain without it.
SpaceX’s AI Campus Plans Put Power, Water And Jobs In Focus
SpaceX outlined two AI computing projects in Texas and Arizona, with local hearings ahead and questions about water, power costs and tax breaks.
How Granite 4.1 3B And 8B Bring LLMON To AI Workflows
IBM says its experimental 3B and 8B models use LLMON to clarify instruction-data boundaries and reduce prompt-injection success on some benchmarks.
What The ChatGPT Voice Retirement Means For Your Account
OpenAI will remove ChatGPT’s Advanced and Standard voice modes on Dec. 11, 2026, moving users to GPT-Live while some features remain on the old system.
Risk-Based Quality Assurance For AI-Coded Bookkeeping
An IdeaNavigator AI proposal outlines a review queue to flag risky AI-coded bookkeeping entries, with a three-firm ledger test suggested for validation.
A Look At Gemini 4 Carbon’s New MKA Test Score And Cost
Google’s Gemini 4 Carbon scores 54.3% on the new MKA benchmark, close to Claude Fable 5.2, at a reported $5.63 per task.
A Second AI Review: GPT-6 Astra Spots Bugs In Claude’s Model Edit
A GPT-6 Astra review found reproducible benchmark and weight-conversion faults in a Claude-led model-edit experiment, while the legal score stayed at zero.