OpenAI introduces GPT-5.2-Codex with stronger security
OpenAI introduces GPT-5.2-Codex, a model that marks another step forward in its internal cybersecurity evaluations. Although it has not yet reached the `High` level, the company is adding safeguards and preparing its systems for future models with greater capabilities.
OpenAI has introduced GPT-5.2-Codex, a new model designed for programming tasks that also raises the capability level observed in its cybersecurity evaluations. The company says this is the third consecutive improvement since GPT-5-Codex.
OpenAI's comparison shows a marked increase in one of its internal cybersecurity evaluations with GPT-5-Codex, another significant jump with GPT-5.1-Codex-Max, and a further improvement with GPT-5.2-Codex.
The company does not publish a specific performance figure in this announcement. It also does not claim that the model has reached the High level of cyber capability defined in its Preparedness Framework. That distinction matters: the model performs better in these tests, but it has not yet crossed the threshold OpenAI uses to classify potential cybersecurity risk as high.
More capability, more controls
OpenAI says it expects future models to follow a similar trajectory. That is why it is designing its preparedness processes as if each new generation could reach the High level, even though this model has not done so yet.
In response to these increased capabilities, GPT-5.2-Codex includes additional safeguards both within the model itself and in the product that provides access to it. The details appear in its security documentation, known as a system card.
In practice, this means progress is not measured only by how well the model writes or understands code. It is also measured by what it could do with instructions involving systems, vulnerabilities, and computer security, and by the limits applied to reduce dangerous uses.
What this means for you
If you use Codex-based programming tools, the announcement points to models capable of handling more complex tasks. One example would be helping review a project, identify problems in code, or work with design prototypes, such as the one OpenAI shows generated by GPT-5.2-Codex.
But the announcement does not support the conclusion that the model is autonomous or can carry out real attacks on its own. The published information describes improvements in evaluations and new protections, not a definitive result in the real world.
The signal worth watching is the distance between model improvements and their security controls. OpenAI is already preparing its systems for a future generation that could reach higher levels of cyber capability, even though GPT-5.2-Codex has not reached that point yet.