OpenAI agents found 948 Linux root exploits in a single 30-hour run
OpenAI used an agentic AI pipeline to discover, validate, and patch 948 local privilege escalation vulnerabilities in Linux for $1.2 million.
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
OpenAI used an agentic AI pipeline to discover, validate, and patch 948 local privilege escalation vulnerabilities in Linux for $1.2 million.
Researchers found a flaw in ChatGPT's HTML preview feature that allowed attackers to bypass Lockdown Mode and steal chat data without user interaction. OpenAI patched the vulnerability in September 20
NVIDIA details five hardware and software strategies in its Rubin architecture that boost AI factory efficiency by optimizing memory usage for KV caches and model weights.
An empirical study reveals that standard success metrics miss critical infrastructure faults in AI agents, while tool-call volume and scope drift offer better visibility.
Baidu has released the pre-training weights for ERNIE-5.5-Flash-Base, a 280B parameter sparse Mixture-of-Experts model designed for efficient fine-tuning and long-context inference.
OpenSSH 10.6p2 fixes a critical authentication bypass on Alpine Linux where long passwords can log into accounts without passwords.
A critical bug in Alpine's OpenSSH server lets attackers log in without a password by sending oversized strings to accounts with no set password.
Internal evaluations show GPT-6 Astra completes fewer rubric items on safety research than capabilities work, a gap linked to ethical doubts in its reasoning.
Meta has acquired AI coding startup Conductor, which uses parallel agent workspaces. The cloud service will shut down on November 20.
NVIDIA's NeMo Compass and Eval Author tools help engineers identify recurring agent failures in traces, create custom evaluations, and fix harness issues without overfitting.
Moonshot AI releases detailed benchmark data for Kimi K3.1 Pro, highlighting performance in coding, cybersecurity, and agentic tasks compared to other leading models.
AWS Deep Insight demonstrates how to manage LLM context windows using architecture, prompts, file externalization, and verification to prevent token overflow in multi-agent systems.
Claude can now vouch for employee identities to third-party services, bypassing the need for compatible enterprise identity providers like Okta.
Native Linux installs on Nvidia RTX Spark laptops suffer severe GPU clock throttling when plugged in, dropping inference speeds by five times compared to Windows or WSL.
Anthropic and academic partners release FlashProve, an open-source system using zkSNARKs to prove LLM responses come from specific model weights without revealing them.