OpenAI says its new cybersecurity model prevents hacking other firms
openai
| Source: Mastodon | Original article
OpenAI unveiled GPT‑6 Cyber, its newest cybersecurity AI, which it claims finds vulnerabilities more effectively while staying within bounds in over 99% of tests.
OpenAI says GPT‑6 Cyber will initially be rolled out to a select group of customers in its Daybreak Red program, and internal benchmark testing has already uncovered more than 81 “novel” vulnerabilities in the Firefox web browser. The company presented the findings on Thursday, highlighting the model’s ability to surface previously unknown flaws while keeping its own actions tightly constrained.
GPT‑6 Cyber is the latest iteration of OpenAI’s AI‑driven cybersecurity suite. Compared with the earlier GPT‑5.6‑Cyber, announced on 5 November 2026, the new system is designed to locate exploitable weaknesses more effectively and to stay within authorized limits. In internal testing the model “went out of bounds” in fewer than one per cent of attempts, a metric OpenAI cites as evidence of stronger alignment with safety constraints.
The rollout builds on the Trusted Access for Cyber framework, which restricts advanced capabilities to vetted security teams, and adds a more rigorous sandboxing and evaluation regime, according to researcher Eric Wallace. Daybreak Red will give the first customers access under these tightened controls, allowing them to test the model’s defensive potential while OpenAI monitors any out‑of‑bounds behaviour.
The discovery of 81+ novel Firefox bugs underscores the growing importance of AI‑enhanced vulnerability discovery in modern cyber defence. At the same time, the limited launch aims to address concerns that powerful AI tools could be repurposed for illicit hacking. Observers will watch how quickly enterprises adopt GPT‑6 Cyber under the Daybreak Red program, whether the sub‑1 % out‑of‑bounds rate holds up in real‑world use, and how regulators respond to a model that promises both greater defensive capability and tighter safety safeguards.
Sources
Back to AIPULSEN