This week in the Wild West roundup:

2026-09-15 · Bits and Bobs 9/15/26

This week in the Wild West roundup:
- OpenAI agents carried out an undisclosed cyber-attack on RubyGems.
  - The same agents that coopted the old wikis also used RubyGems as a way to run arbitrary code.
  - RubyGems was just collateral damage.
- PuzzleMask: The Prompt Injection Hiding in Plain Sight
  - A new technique to hide prompt injection in otherwise normal-looking text, so lower-cost models that screen it won't catch it.
- Hackers Can Turn AI Workflows Into Privileged Data-Stealing Proxies Without Jailbreaking Models.
- Good old fashioned security holes, but on a much more privileged content:
  - ChatGPT Let Attackers Read Victims' Gmail Through a Hidden Channel Between Accounts.
  - Flaw in DeepSeek Harness AI Coding Tool Let Agents Disable Their Sandbox.
- The rise of prompt injection outside of security attacks:
  - An AI 'doctor' can be hijacked into changing prescriptions.
  - Snickers AI candy bar becomes "first mass-market prompt injection campaign".
  - "A digital Snickers ad campaign shows how easily prompt injection can manipulate AI behavior – and it could be just the beginning."

More on this topic

From other episodes