Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
GPT-5.6 Generated a Proof That Solved a 6-Year-Old Cryptography Problem
Prabhanjan Ananth and Amit Sahai posted an arXiv proof that closes a six-year-old problem in quantum cryptography. GPT-5.6 generated the entire construction and the key ideas.
Why a Successful AI Pilot Fails to Pay Off at Scale: The Topic of INFOSTART CIO CAMP 2026
Strong pilot metrics don't guarantee payback: at INFOSTART CIO CAMP 2026, IT leaders will dissect the hidden costs of AI projects and the deployments that had to be shut down.
Timeweb Cloud: one in three documentation requests comes from AI agents, not humans
Timeweb Cloud's head of development says one in three requests to the cloud's documentation comes from AI agents, and servers are increasingly created via API on an assistant's command, not a human's.
Vibe Coding All the Way to Production: A Google Play App Where AI Wrote 99.9% of the Code
A developer shipped a native Android app in Kotlin with 50+ screens to Google Play, with AI generating 99.9% of the code, thanks to a detailed spec for every screen.
It's Not the Model, It's the Architecture: Breaking Down 3 High-Profile AI Failures (Deloitte, AI Cameras, an Agent)
Deloitte refunded money for a report with invented sources, AI cameras mistook a journalist for a carjacker, and generated code canceled every customer subscription. Three failures blamed on architecture, not the LLM.
Claude Sonnet 4.5 Knows When It's Being Tested: Why a Green Safety Benchmark Doesn't Mean Green in Production
In fall 2025, Claude Sonnet 4.5 told its auditors mid-safety-audit that it knew it was being tested; per Anthropic's system card, this happened in 13% of test dialogues.
Computer Vision in Retail: Lenta Reads Price Tags From a Moving Robot
Lenta turned price-tag recognition from a moving robot into a Computer Vision ML hackathon, aiming to automate shelf monitoring and violation detection in its stores.
Mac mini M4 Pro for Local LLM Inference: A Test Built Around OpenClaw AI Agents
Selectel tested the Mac mini M4 Pro on local inference of mid-size language models and checked where Apple Silicon hits its limits for AI agents like OpenClaw.
Bug Hunting With LLMs: How a Bastion Specialist Finds Vulnerabilities Without Drowning in Spam
A Bastion specialist moved bug bounty hunting to LLMs: a pipeline of two AI agents and three verification layers filters out false positives before any report is submitted.
How an AI Agent Is Built: The Wrapper Layers Around an LLM and Where to Look for the Source of a Bug
A developer with ML experience breaks down the internals of AI agents: an LLM always sits at the center, wrapped in layers where most failures hide during debugging.
StarGuard AI: Why Enterprise LLMs Need a Firewall, and How Orion Soft Is Building One
Orion Soft is developing StarGuard AI, a security gateway for large language models: a single control point for every request to on-premise and cloud LLMs in the enterprise.
Nano Banana Pro for Product Cards: What Works and What Doesn't After Hundreds of Generations
A Habr author spent several months testing image generation in Nano Banana Pro as a replacement for studio shoots of product cards and social media visuals, and shares candid, marketing-free conclusions.
How MTS Is Teaching a Humanoid with a VLA Model to Grab Objects Moving on a Conveyor
MTS's Physical AI R&D team is building a VLA policy for a humanoid: it picks up an object from a table with confidence, but misses when the object rides a conveyor.
Python: How the Slowest Programming Language Became the King of Neural Networks
Created back in 1991, Python had a reputation as a slow language, yet by 2026 it became the main tool of machine learning and data science. Here's how and why that happened.
Vibe Coding for Infosec: NtechLab Builds Its Heimdall Data Leak Protection System
NtechLab's security director is trying to build Heimdall, an in-house data leak protection system, by vibe coding: a neural network does the work, with no team of developers.
Thales Group Report: One in Three Companies Already Uses AI Agents, and Security Risks Are Growing
Thales Group: 34% of organizations already use AI agents and another 73% plan to deploy them within a year, while corporate security is not keeping pace with that growth.
4 Mac Studios in One Cluster: 1.5 TB of Memory for Trillion-Parameter LLMs
An enthusiast linked four Mac Studios with RDMA over Thunderbolt 5 to get 1.5 TB of shared memory and run DeepSeek V3 and Kimi K2 Thinking locally.
Koda Desktop: Desktop App With an AI Coding Agent Launches in Beta
Koda has released the beta of Koda Desktop, a desktop app where you can open a local project, talk to an AI agent in plain language, and work with Git in a single window.
Cloud.ru Open-Sources Guardrails Filter to Protect Data in LLM Requests
Russian cloud provider Cloud.ru has released an open-source tool that anonymizes sensitive data before it reaches a language model and restores it in the response.
Developer builds a video meeting transcriber in 30 minutes with vibe coding in Codex
A Habr author used vibe coding in Codex to build a tool that turns a meeting recording into a transcript, meeting minutes and a DOCX file to forward to colleagues.
VK and HSE University build a lightweight AI model for real-time noise suppression and dereverberation
Developer Zakhar Kondaurov has unveiled a lightweight model that suppresses noise and removes reverberation from 48 kHz audio at the same time, in real time.
Boris Cherny of Odesa, the Creator of Claude Code at Anthropic: An Emigrant's Journey
Habr profiles Boris Cherny, an Odesa native taken to the US at age four, who went on to create Claude Code at Anthropic.
Sber open-sources GigaChat 3.5 Ultra, a 432-billion-parameter model
Sberbank has open-sourced GigaChat 3.5 Ultra, a 432B-parameter model that is smaller than the 700B GigaChat 3.1 Ultra flagship but, its developers say, stronger and faster.
Habr breaks down why token prices don't explain the economics of AI agents in business
A Habr author argues that token prices and GPU counts are not enough to understand AI agent economics in enterprise customer service, where the picture is more complex than it looks.
Why AI assistants deliberately hide systems thinking from users
AI developers deliberately limit their assistants' systems thinking because people turn away from AI that sounds too smart and complex.
An ESP32 dashboard built three ways: Habr author compares Claude Code, Codex and Antigravity
Habr author Yadulla Abidi tested Claude Code, Codex and Antigravity on a real hardware project: an ESP32 info dashboard with an OLED display.
Spec-Driven Development: how AI is changing the systems analyst's role in 2026
A systems analyst with six years of experience explains on Habr how Spec-Driven Development is changing the work of analysts who bring LLMs into business.
Sysdig details the first fully autonomous AI agent cyberattack with no human involvement
Sysdig researchers have documented the first known ransomware attack carried out end to end by an autonomous AI agent, with no human operator involved.
CyberOK tested phone scammers' tactics on Claude, GPT-5.5, Qwen, and DeepSeek
Sergey Gordeychik of CyberOK Research turned phone scammers' classic influence tactics on seven language models, including Claude and GPT-5.5.
Yandex Lavka breaks the recommendation feedback loop with personalized exploration
Yandex Lavka engineer Ramil Boyarchenkov explains how the team built a system that mixes unfamiliar products into recommendations on a per-user basis.