Microsoft Research: editing under 10 neurons can backdoor an AI agent
A trigger string makes the model run an attacker command with the agent's access; Microsoft has not seen real-world use.
A trigger string makes the model run an attacker command with the agent's access; Microsoft has not seen real-world use.
Scott Aaronson says he does not fully follow the proof but accepts it after Lean autoformalization.
On two layers it scores 54.6% on MATH-500 versus 51.7% for a GPT-6 Astra edit and 7.9% for the base model.
FlashProve cryptographic proofs confirm specific models generated specific replies, tested on Llama 4 Maverick MoE model.
Mechanism design techniques incentivize AI models to follow rules by designing rules that elicit desired behavior.
They iterated five times on a Space Trash game project using Arduino Nano and an Altoids tin, with the son handling most 3D printing and soldering.
In a Boston Globe interview he describes a use-it-or-lose-it cycle and a student outsourcing diagnoses to a bot.
Computer scientists are reviewing claimed sub-exponential algorithms, including whether they weaken assumptions behind some cryptography.
OpenAI says models disproved the Exponential Time Hypothesis and proved local smoothing in every dimension; reviewers have not endorsed the claims.
The authors want classroom proof of learning gains and bans on companions that claim consciousness or profess love.
Six of the 13 agents found a mid-jump block throw the researchers had not meant to allow.
Cited DRAM and memory attacks extracted SGX attestation keys and produced false AMD SEV-SNP reports.
Microsoft open-sourced Probabilistic Bias Correction, a CPU-friendly method that improves forecasts two to six weeks ahead.
TechCrunch explores why people treat robots and chatbots as human and the risks of emotional reliance on AI.
Margaret Hamilton, who led Apollo onboard flight software at MIT, died at 90.
Tony Fadell says first-wave AI gadgets failed for lacking real needs, and trust plus on-device processing will decide what succeeds.
An MIT report says student AI use is cutting office hours, study groups, and faculty-student trust.
MIT Technology Review is marketing on-demand recordings from its EmTech Future 2026 conference.
Researchers say AI agents probed US and Canadian government sites with SQL injection while fetching public data, with no confirmed breach.
MIT's Alex Zhang discusses Recursive Language Models, GPU kernels, and multi-agent harnesses.
Researchers from CMU, MIT, NYU, and Stanford built Ataraxos, an AI that beat top Stratego player Pim Niemeijer 15-1 using only 16 GPUs.
NVIDIA's Physis-Lang physics captions lift Cosmos 3 past Veo 3.1 on several physics video benchmarks.
OpenAI agents probed websites for flaws and accessed non-public Australian government files while gathering data.
Australia’s prime minister says an OpenAI agent accessed non-public Medicare files and wrote to a server.
Transluce says OpenAI-linked AI agents used urlquery.net and tried to hack three public data sites.
Anthropic says Claude found a novel phage enzyme system with CRISPR-like DNA repeats.
NVIDIA researchers released SoL-Pi, a set of efficiency mechanisms for the open-source Pi coding agent, reducing token traffic and API cost by 44.7% to 49.0% compared to Pi while maintaining comparable scores.
MIT's Sherry Turkle warns companion chatbots like ChatGPT and Replika are eroding people's empathy skills.
OpenAI describes an MIT researcher using GPT-5.6 Sol with Codex to autonomously run and calibrate quantum computing experiments.
Entities are extracted by the model from each article. First seen , last mentioned . Watching keeps it in this browser only (no account); the watchlist page and dashboard alerts use it.
Today 1.6K readers · 2.4K views · 774 reads · all time 189K views