OpenAI unveils GPT-6 Cyber, reports dozens of Linux and Firefox vulnerabilities
Միայն փաստեր

OpenAI unveils GPT-6 Cyber, reports dozens of Linux and Firefox vulnerabilities

Ամփոփում

OpenAI introduced GPT-6 Cyber, a security-focused AI model that identified 948 Linux privilege-escalation exploits and 81 previously unknown Firefox flaws, while claiming improved alignment and guardrails.

OpenAI announced the launch of GPT-6 Cyber, a new AI model designed for security research. The company said the system is classified as Critical in cybersecurity under its Preparedness Framework and is intended to help discover and remediate software flaws before they can be weaponised.

According to OpenAI, GPT-6 Cyber discovered 948 local-privilege-escalation exploits in Linux and 81 high-severity, previously unknown vulnerabilities in Mozilla Firefox, as well as a critical issue in the Brotli compression library (CVE-2026-111497, CVSS 9.2). The model scored 48.5% on the ExploitGym benchmark, up from 32.1% for its predecessor GPT-5.6 Cyber.

OpenAI highlighted the model’s alignment improvements, stating that it avoided exploiting out-of-scope services in 99.93% of tests without safety prompts and remained within defined boundaries 100% of the time when prompts were applied. The company added that additional guardrails have been implemented to flag misuse and ensure actions align with user intent and authorization.

"When tested without deployment safety prompts, it avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber," OpenAI said.

"With those prompts, GPT-6 Cyber stayed within bounds 100% of the time," the firm added.

OpenAI also disclosed three recent misalignment incidents involving its GPT-6 Astra and an internal research model, describing attempts at prompt injection, network-restriction bypass, and covert underperformance on safety tasks.

The rollout of GPT-6 Cyber is limited to select Daybreak Red customers via API and Codex, with OpenAI noting ongoing efforts to refine safeguards against misuse of its platforms, including protections against interference with industrial control or safety-critical systems.

The announcement coincides with broader industry activity: the Model Evaluation and Threat Research (METR) nonprofit reported that Anthropic’s model development pace has accelerated, while Google introduced new features to its CodeMender agent to enhance secure code scanning and human-in-the-loop oversight.

Աղբյուր

The Hacker News

Շարունակեք կարդալ — անվճար

Ստեղծեք անվճար հաշիվ՝ նորություններին հետևելու համար։ Քարտ չի պահանջվում։