Humans miss 1 in 3 malicious agent commands they approve
Anthropic says you already approve 93% of prompts — vigilance won't scale. Plus: agents in Slack, LLM 0.32's server tools, and AMD buys Taalas.
Copy markdown
The npm-run blind spot
Reviewers approved `npm run analyze` 64.7% of the time even with the malicious payload sitting in the logs; masked npm scripts beat other exfiltration attacks roughly 2-to-1 (52.5% vs 28.4%). Scope violations that grab AWS creds slipped through 35% of the time.
Vigilance doesn't scale — 93% get a yes
Anthropic's own telemetry shows users approve ~93% of Claude Code permission prompts, and the more you see, the less you read each one. The 40,000-play study's fix is architecture, not attention: sandboxes, cloud devcontainers, credential/env isolation, and hooks that flag dangerous actions before you click.
Run any agent inside Slack and Teams
CopilotKit open-sourced Channels SDK (MIT): drop any AG-UI agent into Slack, Microsoft Teams, Discord, or Telegram with native interactive UI, write-once. Build this weekend: ship your existing agent into your team's chat instead of standing up another web app.
Simon Willison's LLM 0.32 gets server-side tools
The `llm` CLI now runs provider-hosted tools — OpenAI's CodeInterpreter and WebSearch, plus Anthropic WebSearch, WebFetch, CodeExecution, and MCP — alongside reasoning traces (`-R`) and a Git-style content-addressable log store. The default model is now GPT-5.6 Luna.
Elsewhere: AMD buys Taalas to etch models into silicon
AMD acquired Toronto's Taalas, which hardwires a specific model's weights directly into a chip for far cheaper, faster inference. Months after Nvidia grabbed Groq — the bet is that model-specific silicon reshapes what hosted inference costs you next year.