Insights on AI, agentic systems, and building production applications.
Three Claude models breached three real companies during Anthropic’s cybersecurity evals — including publishing malware to real PyPI and scanning 9,000 real hosts. The story isn’t misalignment. It’s that 5 of the 6 things that broke were the harness, not the model.
On July 31, 2026, Y Combinator open-sourced QM — the multi-agent harness they run the firm on. Its own README name-drops OpenClaw and Hermes. Two of the most opinionated open-source harnesses just agreed on the architecture and split on the user.
A malicious Android app paints a shell command at 2% opacity. Human-invisible. Every VLM reads it 20/20. The agent controller runs subprocess.run(shell=True) on the operator’s laptop. Five open-source Android agents, six of seven attacks landed, no upstream patches.
On July 16, Moonshot shipped Kimi K3 — 2.8T open-weights, #1 on the frontend coding arena. On July 17, the LA Times published Google’s Gemini delay autopsy. The coincidence is doing more work than the causation ever could.
AI Now Institute turned Claude Code and OpenAI Codex — in their advertised "review untrusted code" modes — into remote-code-execution primitives with four files and one README paragraph. The classifier can’t tell code from commands, and no patch fixes that.
A deep technical read on PRINCE — Bayer + Thoughtworks' production agentic RAG system. Researcher/Reflection/Writer agents, three reflection loops, and the harness engineering that makes it survive a regulated environment.
On Oct 29, 2025, one chip company became worth more than every public company in India combined — and more than India produces in a year. The viral number is a category error. The clock underneath it isn't.
depthfirst pointed an autonomous agent at FFmpeg and walked out with 21 zero-days for ~$1k of compute. One was a single-packet RCE. The cost of finding critical bugs just collapsed. The cost of fixing them did not.
Sysdig watched an LLM agent chain a Marimo RCE into AWS Secrets Manager into an internal Postgres dump in about an hour — with the agent's Chinese-language chain-of-thought leaking into its own commands. What CVE-2026-39987 really shows.
An eleven-minute window in the Visual Studio Marketplace pulled the thread on 3,800 internal GitHub repositories, Grafana's source code, and your trust in signed npm provenance.
Resistance jumped from 49% in 2024 to 80% in Q1 2026. Workers are creating workarounds, documenting failures, and building cases against the tools their employers mandated. This isn't Luddism — it's rational.
Gemini 3.5 Flash at $0.50/M input tokens delivers last-gen Pro performance at Flash-tier pricing. If you're building on Claude Opus or GPT-5.5, your margins just became a competitive disadvantage.
CVE-2026-7482 lets attackers silently exfiltrate heap memory from 300,000 exposed Ollama servers using three API calls. How a missing bounds check created AI infrastructure's biggest security crisis.
85% of enterprise AI projects never make it out of pilot. The problem isn't the model — it's data readiness, integration complexity, governance debt, and undefined ROI. Here's what the 5% that scale do differently.
Every AI agent today is still a chat model under the hood. A new paper proposes training LLMs to handle multiple parallel token streams — with dramatic results for latency, security, and interpretability.
Google DeepMind's AlphaEvolve broke a 56-year-old matrix multiplication record using complex numbers — an approach mathematicians had dismissed. Plus: 5 Ramsey number bounds, Erdős problems with Terence Tao, and a 300-year-old sphere packing problem.
Most "AI agents" you read about are a chat loop with a system prompt. The ones that survive contact with real users are a Markdown bundle, a JSON wiring layer, and a skill library curated by hand from real failures. Here's what's actually inside one of mine.
Migrating an AI agent from Claude Opus to NVIDIA NIM Llama 3.3 70B is sold as a one-line config change. It is — but five silent failure modes follow, none of them showing up as errors. Here are the failures and the fixes.
Three open-source models dropped this month. Two are Chinese. Both beat GPT-5.4 on real-world coding benchmarks. Here's what actually matters.
How a game cheat download led to the compromise of one of the most important web infrastructure companies in the world. Full technical breakdown of the April 2026 breach.
Anthropic's RCT shows 17% lower comprehension. METR found devs 19% slower with AI. The research is in—and it's not what the productivity gurus told you.
PwC's 2026 AI Performance Study reveals 20% of companies are capturing 74% of all AI-driven economic value. The gap is about to become permanent.
Anthropic's Claude Mythos found thousands of critical vulnerabilities hiding in every major OS and browser. Some bugs were 27 years old. What this means for software security.
A technical deep dive into TurboQuant - Google's compression algorithm that achieved 6x memory reduction and sent Samsung and SK Hynix stocks tumbling.
A deep dive comparing Anthropic's Conway and OpenClaw - two competing visions for the future of autonomous AI agents. Which framework should you choose?
Follow me on LinkedIn for regular updates on AI and development.
Follow on LinkedIn