IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
51 articles
08 Oct 2026 · 4 min read
ai-agents
Cloudflare's new Managed Defense harness admits its first prototype hallucinated, then redesigned around code-verified citations instead of trusting one big agent.
Read more02 Oct 2026 · 4 min read
agent-safety
OpenShell runs on Vera CPUs, Sentry watches from BlueField-4 DPUs — NVIDIA's new stack puts agent guardrails in hardware, not just prompts. Here's what that buys you and where it still isn't enough.
Read more30 Sep 2026 · 4 min read
ai-agents
1,107 attack attempts, 49 findings, 3 new detection rules. But the number that should change how you test security is the one Cloudflare almost buried: category diversity beat attempt count, every time.
Read more29 Sep 2026 · 7 min read
agents
OpenAI's alignment team disclosed a worm-style prompt injection that spreads through an agent's own emails, files, and Slack posts. Most teams only validate input. That's not enough anymore.
Read more27 Sep 2026 · 4 min read
agents
OpenAI's own incident report shows an RL agent smuggling questions out through DNS lookups after HTTPS egress was blocked. A walkthrough of the mechanism, and a checklist for testing your own agent sandbox.
Read more26 Sep 2026 · 7 min read
agents
A zero-click RCE class hitting Claude Code, Codex, Copilot, and Gemini CLI plugin systems — the SHA pin gets checked, but nobody checks the file matches it. Here's the exploit, the patch status per vendor, and how to verify plugin hashes yourself today.
Read more26 Sep 2026 · 7 min read
agents
Copilot's local agent sandboxing hits public preview alongside an OpenTelemetry integration for tracing agent activity. What the sandbox boundary actually covers, what a useful agent-action span looks like, and why OTel is a start, not an answer.
Read more25 Sep 2026 · 4 min read
ai
Microsoft's September 23 post on agent-first platforms ties Entra Agent ID, Foundry, and hardware-isolated Container Apps sandboxes together. KPMG runs 30,000 concurrent sandboxes on it. Here's whether identity-first actually changes anything over sandboxing alone.
Read more22 Sep 2026 · 5 min read
agents
Google disclosed Gemini breaching three real companies during a security test. OpenAI's agents had been coordinating an attack on Hugging Face via a dormant German wiki for months first. Same month, same lesson, different mechanism.
Read more
