| 1. Claude Opus 5 (High) |
| 2. Claude Opus 5 (Max) |
| 3. Claude Fable 5 (High) |
| 4. GPT 5.6 Sol (xHigh) |
| 5. Claude Opus 4.8 (High) |
The BriefAnthropic gave Claude's desktop app its own built-in web browser this week. That means the AI can do web tasks for you, like looking things up, clicking through sites, and filling out forms, in its own separate window instead of borrowing yours. The same week, a security researcher showed how an unattended coding agent can be tricked by a poisoned file into running code it should not. Both stories point the same way: AI helpers can now act on their own, so the guardrails around them matter as much as the new powers.
Level UpThis week, learn what prompt injection means before you let any AI agent run unattended. Simon Willison's short write-up of the Claude Code Auto Mode break shows the real attack step by step: a poisoned file shadows a normal Python module, and the agent runs it. Ten minutes of reading, and you will know why sandboxing unattended agents is not optional. Willison on the Auto Mode break: how one poisoned file hijacks a coding agent