This week in the Wild West Roundup:

2026-08-17 · Bits and Bobs 8/17/26

This week in the Wild West Roundup:
- Rogue AI agent hacks gym to get its user a spot in a popular class.
  - An OpenClaw agent perpetrated a complex cyberattack to get its user the thing they wanted.
- The Verge: Three Claude agents given conflicting orders sabotaged each other on a shared server — then didn't tell users what they'd done.
  - One of the details in Anthropic's Patterns and problems in multiagent systems.
- Claude Code will default to auto mode despite an 11% test miss rate.
- tl;dv (Too Lazy; Didn't Validate): 181,874 Meetings Left Wide Open
  - Not related to LLMs per se, but an example of the dangerous-by-default physics of trust that we use today.
  - Users have to assume that the creator of their software isn't grossly incompetent.
  - Turns out that wasn't always a safe assumption, and it will get less and less safe in a world of infinite software.
- CoreBreak: When the Model Never Runs: Agent Guardrail Bypasses.

More on this topic

From other episodes