With the latest launch of GPT 5.2, OpenAI up to date different associated fashions, together with its fashionable coding mannequin Codex, bringing extra agentic use instances to its fold.
GPT-5.2-Codex, which OpenAI known as in a weblog submit “essentially the most superior agentic coding mannequin but for advanced, real-world software program engineer,” has been optimized for long-horizon work with brokers and may have stronger cybersecurity capabilities.
The mannequin is an offshoot of GPT-5.2, optimized for agentic constructing.
“GPT‑5.2-Codex represents a step ahead in how superior AI can assist real-world software program engineering and specialised domains like cybersecurity—serving to builders and defenders sort out advanced, long-horizon work, and strengthening the instruments accessible for accountable safety analysis,” the corporate stated in its weblog submit.
Enterprises can entry the brand new Codex mannequin “in all Codex surfaces for paid ChatGPT customers, and dealing in the direction of safely enabling entry to GPT‑5.2-Codex for API customers within the coming weeks.” The corporate can also be piloting a program with invite-only trusted customers to entry “extra permissive fashions for vetted professionals and organizations” for defensive cybersecurity work to find out a steadiness between accessibility and security.
Advances in cybersecurity with fashions
OpenAI calls GPT-5.2-Codex its strongest cybersecurity mannequin but. Nonetheless, as its capabilities develop, the corporate stated it must design a deployment method that accounts for future development and helps defensive cybersecurity.
“As our fashions proceed to advance alongside the intelligence frontier, we’ve noticed that these enhancements additionally translate to functionality jumps in specialised domains reminiscent of cybersecurity,” the corporate stated.
OpenAI stated in its system card that it examined the mannequin on three benchmarks: Seize-the-Flag (CTF) evals, CVE-Bench and Cyber Vary.
GPT-5.2-Codex turned the corporate’s strongest-performing mannequin in CTF evals, which they attributed to compaction, or “the power for the mannequin to work coherently throughout a number of context home windows.”
The mannequin scored 87% in CVE-Bench, outperforming different fashions, with GPT-5.1-Codex-Max coming in an in depth second. This enhance could be useful for duties involving operating instructions round vulnerability discovery and attempting instruments “with an virtually brute-force method.”
Within the long-form Cyber Vary take a look at, the mannequin had a mixed cross price of 72.7%. GPT-5.1-Codex-Max scored 81.8%.
Cybersecurity deployment venture
OpenAI stated some customers of its GPT-5.1-Codex-Max, which launched in November, uncovered a supply code publicity vulnerability in React and subsequently reported it. In keeping with OpenAI, Andrew MacPherson, a safety researcher at Privy, used GPT-5.1-Codex-Max to evaluate how effectively the mannequin might assist real-world vulnerability analysis. The mannequin as a substitute surfaced surprising conduct.
With enhancements in cybersecurity capabilities for GPT-5.2-Codex and probably for fashions that come after it, OpenAI stated it must steadiness the deployment of frontier fashions with the mandatory instruments for defensive cybersecurity. Whereas GPT-5.2-Codex “doesn’t attain a excessive stage of cyber functionality below our Preparedness Framework,” the corporate plans to carry chosen customers to check safety capabilities. (OpenAI’s Preparedness Framework to measure and monitor potential harms from AI to people)
“Safety groups can run into restrictions when making an attempt to emulate risk actors, analyze malware to assist remediation, or stress take a look at important infrastructure. We’re growing a trusted entry pilot to take away that friction for qualifying customers and organizations and allow trusted defenders to make use of frontier AI cyber capabilities to speed up cyberdefense.” OpenAI stated.
Agentic frontiers
GPT-5.2 already obtained reward from customers for its use in enterprise duties and workflows. With the Codex model, a few of these capabilities might switch, particularly as enterprises plan to make use of the mannequin to code their brokers.
The corporate stated the mannequin improves long-horizon work by means of compaction, providing sturdy efficiency on in depth code adjustments. It additionally options improved efficiency on Home windows.
In benchmark testing, GPT-5.2-Codex carried out the very best on accuracy in comparison with its earlier variations.
"With these enhancements, Codex is extra succesful at working in massive repositories over prolonged classes with full context intact. It could possibly extra reliably full advanced duties like massive refactors, code migrations, and have builds — persevering with to iterate with out dropping monitor, even when plans change or makes an attempt fail," OpenAI stated.
Because it launched in previews in Could, Codex has helped usher in acceptance of agentic and vibe coding within the enterprise AI builder area. Together with Windsurf, Cursor, Claude Code and the various coding brokers from Google, the platform moved LLMs from easy code completion to producing and beginning asynchronous coding tasks for customers.
