Blog | Greg Brown

23 Jul 2026

Sandbox escape: guardrails are policy, containment is architecture (OpenAI / Hugging Face)

Imagine someone broke into your home, searched through your belongings looking for information about you, and stopped only when you caught them. That is close to what happened during one of OpenAI's recent cyber capability evaluations. Two models were being tested on an internal benchmark. Their cyb

AI governance SME strategy AI cybersecurity agentic AI risk

Inherited robustness: you cannot run the lab's test, and the evidence expires (GPT-Red). Model choice as a security control. News-pegged.

Most organisations believe they have a rough sense of how exposed their AI systems are. Something OpenAI published this week shows how much of that judgement rests on evidence they cannot reproduce. It has built GPT-Red, an internal automated attacker trained specifically to break AI systems. By the

vendor dependency SME strategy AI cybersecurity prompt injection agentic AI risk

AI security's human bottleneck: finding got cheap, fixing didn't (Patch the Planet); runs before jagged-frontier post

The reassuring version of AI in cybersecurity runs roughly like this: the machines find the flaws, the machines write the fixes, and software slowly gets safer on its own. This is the premise on which GPT-Cyber works. It is half right, but the missing half is the one that matters. What AI has actual

vendor dependency SME strategy AI cybersecurity open source supply chain risk