GPT-5.3-Codex crossed a threshold in February 2026: it became the first AI model in OpenAI's history to receive a "high" rating on their cybersecurity preparedness framework — meaning meaningfully elevated capability to assist with offensive security tasks.
What "High" Cybersecurity Risk Means
OpenAI's Safety Response
What Development Teams Must Do Now
- Audit API keys and secrets in all repositories
- Review authentication and authorisation logic targeting privilege escalation paths
- Update dependency inventories — vulnerable libraries are easier to exploit at scale
- Implement structured logging of all privileged actions
- Brief teams on social engineering escalation — AI-generated phishing is harder to detect
Our custom software development and cloud and DevOps practices incorporate security review as a standard phase.
How we pick topics, use AI, check claims and correct mistakes: our editorial policy.
