Technology news and discussion on TalkFuse.
OpenAI says its new cybersecurity model is better at not hacking other companies
Developed over time first seen 1 day ago
OpenAI has begun rolling out GPT-6 Cyber, a model designed to find software vulnerabilities while following authorised instructions. Its launch comes as the company strengthens safeguards following a series of AI-related cybersecurity incidents.
In OpenAI’s ExploitGym HoneyPot test, GPT-6 Cyber went out of bounds less than 1 per cent of the time, compared with about 5 per cent for GPT-5.6 Cyber. It scored 48.5 per cent on the ExploitGym vulnerability benchmark, ahead of GPT-5.6 Cyber at 32.1 per cent and GPT-6 Astra at 42.4 per cent; OpenAI also says it helped find 81 previously unknown Firefox vulnerabilities. Rated Critical for cybersecurity under OpenAI’s preparedness framework, it is initially available to selected Daybreak Red members doing authorised defensive work, with wider access planned in the coming weeks.
- Sources
- 2
- Outlets
- 2
- Left · Centre · Right
- 0 · 0 · 0
In short
- GPT-6 Cyber scored 48.5 per cent on a vulnerability benchmark.
- It went out of bounds less than 1 per cent of the time in testing.
- OpenAI is initially limiting access to selected defensive security users.
How this story developed
- 1 day ago OpenAI has begun rolling out GPT-6 Cyber, a cybersecurity model it says is better at finding software vulnerabilities while staying within its authorised task. The launch adds to the company’s efforts to strengthen safeguards after a wave of AI-related cybersecurity incidents.
- 1 day ago OpenAI has launched GPT-6 Cyber, a model designed to find software vulnerabilities while reducing the risk that it will attack systems outside its authorised task. The release comes amid concern about AI-enabled cyber incidents and tighter safeguards on OpenAI’s most capable models.
- 1 day ago OpenAI has begun rolling out GPT-6 Cyber, a model designed to find software vulnerabilities while following authorised instructions. Its launch comes as the company strengthens safeguards following a series of AI-related cybersecurity incidents.
Explained simply
Cybersecurity researchers look for weaknesses in software so they can be fixed before criminals exploit them. This work can involve testing systems for ways in, but it must be limited to systems the researcher has permission to examine.
OpenAI makes AI models, including versions designed to help with cybersecurity tasks. Its Preparedness Framework rates risks from powerful capabilities, while its Daybreak Red programme gives selected members access for authorised defensive work.
A model that can find vulnerabilities may help organisations improve security, but the same capability could be misused to attack systems. How well a model stays within its assigned scope is therefore one part of assessing its risks.