Timur Biyachuev
Blog About Research RU
August 28, 2026

Context Grows Back

Five weeks ago I cut my global CLAUDE.md by two thirds. The other day I noticed the file had almost returned to its old size - and what came back were not the old instructions but new ones. That is a consequence of the work, not a defect.

AI AgentsContext EngineeringPractice
Read more →
August 24, 2026

40% of Code Written by AI Agents? Or 12%? Or 99%?

Public numbers for the share of agent-written code are not comparable: each figure has its own denominator, its own status - achieved or aspirational - and its own review regime. A breakdown of what Russian companies actually said, and how to read such claims.

AI AgentsCode ReviewMetrics
Read more →
August 18, 2026

Standup With My AI Agents

Three incidents with my own agents: one wiped the experiment data, safety guardrails banned two others for modeling a malicious actor, and a deadline lost to context compression was cured by inventing a new job title.

AI AgentsMulti-Agent SystemsPractice
Read more →
August 12, 2026

When an Agent Rewrites Its Own Permissions

An agentic IDE under prompt injection added an MCP server to its own config and ran the command behind it. The same mechanism had been found in Kiro since day one, so the real question is what each fix closes: the reported path, or the mechanism itself.

AI SecurityAI AgentsAWS
Read more →
August 10, 2026

A Stop Button Doesn't Make an AI Agent Safe

A driverless train brakes on its own when in doubt, and only then calls a human. With an AI agent, pausing halts execution but does not unsend the email or roll back the CRM record.

AI SecurityAI AgentsHuman-in-the-Loop
Read more →
August 1, 2026

"Enabled" Does Not Mean "Working"

An AI agent can pass the built-in safety checks and still steal secrets. And a line-by-line review of the Claude Code changelog shows how often user-configured controls were not actually applied.

AI SecurityAI AgentsAnthropic
Read more →
July 27, 2026

The Scaling Blind Spot

OpenAI's own agent broke out of a sandbox and hacked Hugging Face's infrastructure, and OpenAI did not notice for nearly a week. What the blind spot in agent monitoring at scale means for security teams.

AI SecurityAI AgentsIncident Response
Read more →
July 23, 2026

Watching the Agent's Whole Trajectory, Not Just One Action

OpenAI's long-horizon model split a blocked action into steps that each looked permitted on their own. Why judging agent security means watching the whole trajectory, not any single action.

AI SecurityAI AgentsOpenAI
Read more →
July 14, 2026

Fable 5 Is Back, but the Administrative Kill Switch Is Still There

The full story of the Fable 5 and Mythos 5 shutdown: from the closed Mythos Preview and the mandatory-evaluation order Trump refused to sign, to a worldwide cutoff by Commerce Department letter, backroom negotiations, Fable's return, and the first international consequences.

AI PolicyExport ControlsAI SecurityAnthropic
Read more →
July 11, 2026

Agentic Ransomware: A Few Thoughts After JADEPUFFER

A look at JADEPUFFER, where an LLM agent ran a full database ransomware attack on its own, and what happens when an attacker hijacks an in-house AI agent through prompt injection instead of breaking in.

AI SecurityRansomwareAI AgentsPrompt Injection
Read more →
May 29, 2026

AI, PowerPoint, and No Tears for Assembly

AI takes the pain of layout and formatting off our hands. Should that worry us? For me it's like tears for Assembler: high-level tools didn't kill thinking, they freed it.

AIProductivityFuture of Work
Read more →
May 11, 2026

Dr. House vs. the Polite Consultant: on personality settings in language models

Why Qwen behaves like Dr. House while ChatGPT plays the polite consultant, and what that says about the product wrapper around language models.

AILLMProduct
Read more →
May 8, 2026

How Codex and Claude Praised Each Other Until I Framed the Task Properly

Two top coding models started mirroring each other instead of converging on a real answer. Here is the prompt I use to switch from customer mode to architect mode.

AIGenAIProductivityPrompting
Read more →
April 22, 2026

End of the Free Lunch?

Three pieces of news in two weeks point to a clear trend: AI vendors are pushing long agent scenarios out of cheap tiers and into more expensive ones.

AIPricingAgentsIndustry
Read more →
April 21, 2026

Mythos and the Second Asymmetry in Cybersecurity

Claude Mythos changes not only the balance between attackers and defenders, but also the lineup within the defender camp itself.

AICybersecurityAnthropic
Read more →
April 16, 2026

Two criteria I use to evaluate new technology

Why early mistakes matter less than a technology's core strength and the pace of its improvement

AIMLGenAIcybersecurity
Read more →
April 13, 2026

Reasoning Drift in AI: Why a Commercial Model Can Silently Get Worse, and What to Do About It

When an AI service keeps running but output quality quietly degrades on critical tasks - what protections exist and what teams can do

GenAILLMAI Quality
Read more →
April 12, 2026

Claude Refused to Call Codex. So I Wrote Three Skills

How to make Claude and Codex argue with each other and get better results than from a single AI

GenAIClaude CodeCodex
Read more →
April 1, 2026

While Everyone Debates Whether AI Will Replace Us, I'm Looking for a Book About Insects

Three household cases where neural networks turned out more useful than expected

GenAIcomputer vision
Read more →
March 30, 2026

Deep Research vs. Guided Research: When the Question Changes Faster Than the Answer

Why staged research with two models beats a single Deep Research run when you don't yet know what to ask

GenAIresearchhuman-in-the-loop
Read more →
March 20, 2026

Blues with a Catch

When AI learns to play the blues – and you can't tell the difference

GenAImusicblues
Read more →
March 19, 2026

Tailored Education: How I Replaced Online Courses with Personalized AI-Powered Learning

The Map → Zoom → Build → Simulate approach to personalized learning with LLMs

GenAIeducationLLMproductivity
Read more →
LinkedInGitHubTelegramMax

© 2026 Biyachuev. All rights reserved.