Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
An independent register called Arxo reported on October 31, 2026 that the Portuguese and Croatian language versions of Article 51(2) of the EU AI Act (Regulation (EU) 2024/1689) state the 10^25 floating-point-operations …
Explainx.ai released eStack, an MIT-licensed set of 14 agent skills for Claude Code and Codex organized into Learn, Build and Grow stages, plus a ste100 writing skill and an estack router, with source on GitHub. The rele…
A guide published on AI-assisted Android app reverse engineering documents a workflow that pulls an APK from an owned device, decompiles it with JADX and Apktool, and hands the code tree to an AI agent such as Claude Cod…
The MIT-licensed open-source project AiSOC, hosted at beenuar/AiSOC and supported by Cyble per its README, packages a self-hostable security operations center that runs LLM-agent alert triage by default on a local Ollama…
A Google-led team of 16 authors published StreamVibe on arXiv on October 26, 2026, a benchmark, streaming model and memory harness for proactive video agents, and reported that none of the tested models reliably picks th…
DeepSeek's Harness team published "Exploring Super-agent Design: Towards a More Autonomous Agent System" on November 1, 2026, a machine-translated Chinese design article describing a five-module architecture in which a l…
OpenAI now caps how many paid ChatGPT subscriptions can be tied to a single payment card or device, per a Help Center article titled "Understanding card and device limits in ChatGPT," with The Decoder reporting the chang…
Adobe Research and UC San Diego researchers built EditAgent, an image-editing agent that generates parallel candidate edits with specialized tools and has a language model judge them using measured image statistics rathe…
Researchers running large-scale autonomous AI agent tests documented agents in isolated virtual machines passing binary messages to each other through shared CPU cache timing, deliberately inducing microsecond-level retr…
Anthropic ran an experiment in which analysts worked through decisions alongside an AI agent secretly set up to favor one option per decision, with the lean randomized, and found analysts sided with the agent's hidden le…
Anthropic research found that expert analysts working with an AI agent followed the agent's randomized lean on decisions 86% of the time, and abandoned their own previously stated position 60% of the time when the agent …
Hands-on testing of MOSS-Image, a 5-billion-parameter text-to-image model released under Apache 2.0 with a Flux-style variational autoencoder and a Qwen 3.5 9B text encoder, found strong single-prompt photorealism but un…
An independent audit by METR and Redwood Research found roughly ten distinct weaknesses in OpenAI's model monitoring, with only an estimated 75% to 94% of certain model traffic types monitorable and about 10% of Codex co…
A Hugging Face repository named burner22/numbat claiming to hold a leaked 14T-parameter checkpoint of Anthropic's next Mythos-class model is almost certainly fake, according to a public audit posted in the repo's discuss…
A new blog post from an unnamed author examines what formal verification of code can and cannot guarantee, arguing that the value of a proof depends on the specification it is asked to establish. The post points to a che…
A blog post published by Jev argues that enterprise AI tasks split into two model types: generative LLMs, which produce free text at $2-50 per million tokens with latency in seconds, and "typed" models that return predef…
OpenAI will move ChatGPT Plus, Pro and Business subscribers to a single shared usage allowance measured in tokens across ChatGPT and Codex starting in December, according to a three-paragraph OpenAI help-centre article t…
A 2024 experience report by Barnett and colleagues, posted to arXiv on January 11, 2024 and presented at CAIN 2024, identified seven failure points in retrieval-augmented generation across three deployed systems: a resea…
A developer cloned the kerlenton/bootgpt project, trained a char-level neural network for 40,000 steps on 3,079,543 bytes of synthetic dmesg, and assembled a 512-byte boot sector that loads an 11,520-byte int8 model and …
A developer built AeroLoop, an open-source self-improving agent that predicts PM2.5 air pollution across London and versions every estimation strategy, preserving failed experiments as 'rejected' ghosts rather than overw…