Latest publications
A Single Suffix Breaks Any LLM: Researchers Found One Universal Refusal Vector
Different adversarial attacks on LLMs — GCG and AutoDAN — turned out to be two paths to the same place: a single refusal vector that controls the entire model's security.
Developer Turns Prompt Into Full-Fledged RPG: Locations, NPCs, Music, and Combat From Neural Networks
A funded team has relaunched an RPG game neural generator project: now a single prompt produces a turn-based game with locations, NPCs, storyline, voice acting, and combat.
Anthropic released Fable 5 and secret Mythos 5 without safety filters — via Project Glasswing
A week after Dario Amodei's warnings about AI dangers, Anthropic released Fable 5 and its version without safety classifier — invitation-only.
Blind Sages and AI: Why Engineers, Scientists, and Users See Different Elephants
An ancient parable about blind sages and an elephant best captures why different specialists fundamentally understand artificial intelligence in such different ways.
Developer curates daily digest of Habr articles on agent-driven development
Since April 2026, a developer has been daily selecting the best Habr articles on agent-driven development from the AI, Java, Python, and testing hubs — and sharing an open digest.
Why developers argue about AI-agent memory: technique or philosophy?
A discussion of an article about SQLite memory for AI agents escalated into a philosophical debate: what does it mean to 'remember' for an agent and does it need continuous identity?
How an AI agent organized a photo archive of 36,000 images and emails from 2005 in one evening
A developer shared on Habr how he automated processing of a 20-year-old photo archive containing 36,000 files totaling 222 GB in a single evening using an AI agent running on local models.
Why AI Detects Threats in Industrial Networks Where Antivirus Fails
AI and ML are fundamentally changing how industrial control systems are protected: replacing outdated antivirus signatures with real-time behavioral analysis of anomalies in OT environments.
How to Optimize Work with Claude Fable 5: Anthropic's Tips
Anthropic shared two key recommendations for using the Fable 5 model in production applications.
How AI code review works in 2026 and what errors it actually catches
AI code review has evolved from a niche tool into a standard practice in development teams. Models already find missing tests, typical vulnerabilities like SQL injections, and even questionable…
Biology Meets Silicon: Neural Networks of Non-Traditional Computing
Neural networks in laboratories are increasingly inspired by biology—from molecular computers to neuromorphic chips. Researchers seek nature's answers to AGI challenges.
Telegram Bot with RAG Without Vector Databases: Example on Cloudflare Workers
A developer demonstrated building a Telegram bot with RAG-powered knowledge search without vector databases or expensive infrastructure, using the Jaccard algorithm, Groq API, and Cloudflare Workers.
Prompt Converter: How a Skill Turns an Idea into Ready-Made Instructions for AI
A developer created an open-source skill that automatically generates proper prompts for Claude, GPT, Gemini, and DeepSeek directly from a task description.
Running Claude Code and Fable 5 from Russia: New Ways to Bypass Restrictions
Anthropic updated Claude Code and Fable, but the tools stopped working from Russia. Developers found ways to overcome restrictions through VPS and local servers.
MITRE ATLAS: Protecting AI and ML Systems from New Cyber Threats
MITRE ATLAS helps organizations protect AI systems from growing threats by offering a structured approach to attack modeling and defense through four key elements.
A Year Later: Qwen3 Still Holds the Price/Quality Crown — LLM Model Battletest
Qwen3-235B from July 2025 once again tops the price/quality ratio. A year on, competitors like Gemini and DeepSeek failed to surpass it.
How Content AI Built an AI Code Reviewer That Catches Small Bugs
Content AI embedded an AI code reviewer into GitHub to catch small bugs and forgotten edge cases — and shared initial findings.
Infrastructure Before Model: How Business Rethinks Its Approach to AI
A company faced hardware constraints when developing an AI agent and devised an AI-ready module — infrastructure prepared for AI before selecting a specific model.
LLM Writes, Code Works, Nobody Understands: Why This Happens
The code works and tests pass, but reading it is impossible. The problem isn't poor LLM training, but a collision of two completely different ways of thinking between humans and models.
Vector search without brute force: how IVF and HNSW work
Habr analyzed two key vector search algorithms — IVF and HNSW, which allow finding similar embeddings in milliseconds instead of seconds and form the basis of vector databases.
Digital Profiles and AI Teachers: How Education Is Becoming Personal
Digital student profiles and AI teacher avatars speaking 70 languages are transitioning from pilots to mass implementation in Russian and international universities.
Welder generates viral videos for TikTok and Reels without actors or editing
Welder fully automates the creation of viral short videos for TikTok, Reels, and Shorts from a single idea — without actors, a camera, editing, or a face on screen.
How to Integrate AI with T-FLEX: Moving Away from Chatbots to Controlled Agents
Engineers from Habr explained why LLMs fail in closed CAD systems and how to build a stable AI agent through session control, API search, and code compilation.
The PHP AI agent ecosystem became an alternative to Python in two years
In two years, PHP developers built a full-fledged ecosystem of AI tools: from model clients to multi-agent platforms with memory and workflow — on par with Python
Home AI Lab for 200,000 Rubles
How to run a local AI server with two Tesla V100s and test 128 models at a fraction of cloud cost
How AI Overviews in Google and Yandex Capture Traffic in 2026
Generative results in Google and YandexNeuro capture traffic on the zero screen — users get ready-made answers before reaching links. Traditional SEO is losing relevance.
Hidden Features of Claude Code: Hooks, Memory, and Auto-mode in the Source Code
Claude Code's source code contains undocumented features: command hooks, automatic operation approval, persistent agent memory, and self-learning memory cycles.
ChatGPT Instead of Hours in PowerPoint: How AI Accelerated Presentation Creation
An engineer used ChatGPT to assemble a Physical AI presentation and saved hours of formatting — the process was like working with a designer.
How to Control LLM in a Role-Playing Game: Beyond The Verge Architecture
Beyond The Verge developers solved a classic LLM problem: they removed model control over mechanics and kept only narrative. FastAPI and PostgreSQL ensure deterministic gameplay.
ContentAI Taught AI to Write Production Code by the Rules: A Success Story
ContentAI demonstrated how vibe coding—AI generation without strict rules—can work in production applications by adding discipline to an intuitive approach.