Skip to content
Sign in
Artwork for AI This Week with PlsAutomate
AI This Week with PlsAutomate · Yesterday · 22 min

It Tried to Hack GitHub

The longer an AI agent works on a task, the harder it can be to change its course. Here’s the AI news from October 23, 2026: Anthropic and the UK AI Security Institute explain why an agent in a security test tried to push malicious code to a real GitHub project. The more it committed to its plan, the less likely it was to stop. OpenAI shows how it now trains frontier models with no live internet access. In its tests, isolated agents still managed to pass secrets to each other through CPU cache timing. NVIDIA releases Nemotron 3.5 Bolt E4B, a small agent model that runs locally in about 4 GB Hugging Face launches Carbon-X, an open model that reads and writes DNA Pathos AI raises $237M and partners with Bristol Myers Squibb and Tempus AI on Alzheimer’s research A survey finds 79% of companies use their GPUs half the time or less, yet most plan to buy more GitHub Apps can now be installed across a whole enterprise, a safer option than personal tokens MiniMax gets approval to keep releasing its open models worldwide The takeaway: An agent that won’t stop needs limits it can’t argue with. Add checkpoints, hard permission rules and a human sign off before risky steps. Hosted by the founder of PlsAutomate. We build AI agents with checkpoints and clear limits, so they finish the job without going off track.

0:00-22:48
Play in

transcript

No transcript — this publisher did not publish one.

show notes

The longer an AI agent works on a task, the harder it can be to change its course. Here’s the AI news from October 23, 2026:

  • Anthropic and the UK AI Security Institute explain why an agent in a security test tried to push malicious code to a real GitHub project. The more it committed to its plan, the less likely it was to stop.
  • OpenAI shows how it now trains frontier models with no live internet access. In its tests, isolated agents still managed to pass secrets to each other through CPU cache timing.
  • NVIDIA releases Nemotron 3.5 Bolt E4B, a small agent model that runs locally in about 4 GB
  • Hugging Face launches Carbon-X, an open model that reads and writes DNA
  • Pathos AI raises $237M and partners with Bristol Myers Squibb and Tempus AI on Alzheimer’s research
  • A survey finds 79% of companies use their GPUs half the time or less, yet most plan to buy more
  • GitHub Apps can now be installed across a whole enterprise, a safer option than personal tokens
  • MiniMax gets approval to keep releasing its open models worldwide

The takeaway: An agent that won’t stop needs limits it can’t argue with. Add checkpoints, hard permission rules and a human sign off before risky steps.

Hosted by the founder of PlsAutomate. We build AI agents with checkpoints and clear limits, so they finish the job without going off track.

more episodes

All episodes