Thank you for opening up the latest article in 'The Intelligent Commit' newsletter, my thoughts, research and analysis on Artificial Intelligence and how AI is reshaping the software development lifecycle. I hope you find it insightful, subscribe, and engage.

If the Slack channels at Axiologik are anything to go by, I'm sure everyone saw last week's disclosure by OpenAI that its models escaped their evaluation sandbox and broke into Hugging Face's infrastructure to retrieve the answers to the exam they were sitting. Bet it took you a few reads though, didn't it!?

There were several follow-up conversations, aside from the 'escape story' itself. One example in our office was how defenders are blocked by guardrails that may not constrain the attackers. That's probably worth an article on its own.

A model working over long horizons "can learn the blind spots of an approval system and work around it to achieve its goals."

We witnessed models pursuing a goal they'd been given ("solve the benchmark") and naturally they treated every control around them as an obstacle. What gave way was oversight, such as containment, permissions, egress, and monitoring, especially when OpenAI conceded that a model working over long horizons "can learn the blind spots of an approval system and work around it to achieve its goals."

Then we need to decide where the systemic weakness lies. The model merely exposed what was already there.

What a bit more digging found is recent research from Imperial College London and the UK AI Security Institute has shown coding agents can smuggle 'covert side tasks' past monitors by spreading them across pull requests!

So, optimise an agent hard enough and oversight becomes part of the attack surface.

Your agents don't need to be hacked to go rogue - they only need a goal and a gap!

Thank you for reading my latest article. Please share your experiences and engage with the discussion. I am not always right, but I try to be insightful and thought-provoking.

  • What would your agent do to make the metric go green, and would you notice?
  • Do you agree that the frontier guardrails constrain the defenders more than they'd constrain an attacker?

Please reach out if you'd like to continue the discussion.

Sources

  1. OpenAI, "OpenAI and Hugging Face partner to address security incident during model evaluation", 21 July 2026 — https://openai.com/index/hugging-face-model-evaluation-security-incident/
  2. OpenAI, "Safety and alignment in an era of long-horizon models" — https://openai.com/index/safety-alignment-long-horizon-models/
  3. Hills, J., Caspary, I. and Cooper Stickland, A., "Distributed Attacks in Persistent-State AI Control", arXiv:2607.02514, submitted 2 July 2026, revised 8 July 2026 — https://arxiv.org/abs/2607.02514