OpenAI on Tuesday unveiled GPT-5.4-Cyber, a variant of its newest flagship mannequin, GPT‑5.4, that is particularly optimized for defensive cybersecurity use circumstances, days after rival Anthropic unveiled its personal frontier mannequin, Mythos.
“The progressive use of AI accelerates defenders – these answerable for protecting methods, knowledge, and customers protected – enabling them to search out and repair issues sooner within the digital infrastructure everybody depends on,” OpenAI stated.
At the side of the announcement, the unreal intelligence (AI) firm stated it is ramping up its Trusted Entry for Cyber (TAC) program to hundreds of authenticated particular person defenders and tons of of groups answerable for securing crucial software program.
AI methods are inherently dual-use, as unhealthy actors can repurpose applied sciences developed for reliable functions to their very own benefit and obtain malicious targets. One core space of concern is that adversaries might invert the fashions fine-tuned for software program protection to detect and exploit vulnerabilities in widely-used software program earlier than they could be patched, exposing customers to vital dangers.
OpenAI stated the aim is to democratize entry to its fashions whereas minimizing such misuse, as properly as strengthening its safeguards by means of a deliberate, iterative rollout. The thought is to allow accountable use at scale, give defenders a head begin, and concurrently shore up guardrails towards jailbreaks and adversarial immediate injections as mannequin capabilities develop into extra superior.
“As mannequin capabilities advance, our strategy is to scale cyber protection in lockstep: broadening entry for reliable defenders whereas persevering with to strengthen safeguards,” the corporate added.
The ChatGPT maker, which launched Codex Safety as a strategy to discover, validate, and suggest fixes for vulnerabilities, revealed that the AI-powered utility security agent has contributed to over 3,000 crucial and excessive mounted vulnerabilities.
OpenAI’s restricted launch follows the preview of Anthropic’s Mythos, a frontier mannequin that is being deployed in a managed method as a part of Challenge Glasswing. The mannequin, the corporate stated, discovered “hundreds” of vulnerabilities in working methods, internet browsers, and different software program.
“The strongest ecosystem is one which constantly identifies, validates, and fixes security points as software program is written,” OpenAI stated. “By integrating superior coding fashions and agentic capabilities into developer workflows, we may give builders quick, actionable suggestions whereas they’re constructing, shifting security from episodic audits and static bug inventories to ongoing, tangible danger discount.”
