The AI agent story keeps splitting into two tracks: better workflows and bigger attack surfaces. Anthropic's Claude Opus 4.8 system card and a wave of posts on 'build agents not pipelines' and permission bottlenecks show engineers rethinking how to structure agentic coding, while a parallel wave of security research (fake Anthropic phishing sites, DOM-exfiltration via browser extensions, C2 channels hiding in AgentCore's S3 usage) shows attackers targeting that exact same tooling. On the product side, MiniMax M3 landed on Ollama Cloud, a ternary 1-bit image model shipped for local devices, and DigitalOcean shipped intelligent model routing for OpenCode, all pointing toward cheaper, more distributed inference. Meanwhile NixOS 26.05 dropped and a couple of deep systems posts on backpressure and diffing are worth a slow read. Funding news underscores where VC attention is heading: $50M for an AI that picks good research questions rather than answering them faster.

• Anthropic published the full Claude Opus 4.8 system card
• Permissions, not model quality, are the real agent bottleneck
• Fake Anthropic sites are pushing fileless infostealers at Claude Code users
• MiniMax M3 now runs on Ollama's Cloud
• A ternary 1-bit image model shipped for on-device generation