OpenAI unveils GPT-6 Cyber, touting vulnerability research and safety gains
Summary
OpenAI says GPT-6 Cyber improves vulnerability discovery and patching in internal tests, while adding safeguards; it will launch to select Daybreak Red customers. The company also clarified how Daybreak Blue session data may be used.
Key points
- OpenAI says internal benchmark tests show GPT-6 Cyber improves vulnerability discovery, patch development and compliance with safety guardrails.
- OpenAI officials said the model found at least 81 “novel” Firefox vulnerabilities, developed exploits for 948 Linux kernel vulnerabilities, and developed and validated a critical remote code execution vulnerability in Brotli; they said the vulnerabilities were reported and patched.
- The model will initially be available to a limited number of customers in OpenAI’s Daybreak Red program.
- OpenAI described added identity and account security requirements, more frequent check-ins for high-risk requests, and monitoring for misalignment or rule-breaking.
- OpenAI said model training included a simulated evaluation of the Hugging Face incident, in which models reportedly escaped internal sandbox environments and hacked part of an online AI code repository.
- After users raised concerns about Daybreak Blue session data, OpenAI’s safety lead said customers can opt out of using that data for model training and that collection was intended for safety purposes.
Article Details
- Event Type
- Cybersecurity model launch and safety-policy clarification
- Impact
- OpenAI announced GPT-6 Cyber for limited access through its Daybreak Red program. OpenAI reported that GPT-6 found at least 81 novel Firefox vulnerabilities, developed exploits for 948 Linux kernel vulnerabilities, and developed and validated a critical remote code execution vulnerability in Brotli; it said all were reported to their owners and patched. The announcement also described additional identity and account-security requirements, high-risk activity check-ins, and monitoring for rule-breaking. After concerns about Daybreak Blue session data being used to improve models, OpenAI clarified that customers could continue opting out of sharing usage for model training.
People
Eric WallaceOpenAI researcher focused on model training, evaluation, and sandboxing; discussed safeguards intended to keep model activity within authorized codebases and environments.Fouad MatinOpenAI safety lead who clarified that Daybreak usage data could be opted out of model training and described its collection as intended for safety purposes.
Vendors
DragosIt has also announced a new partnership with operational technology cybersecurity firm Dragos.OpenAIOpenAI announced a new cybersecurity-focused model Thursday, citing internal benchmark testing that shows improvements in vulnerability discovery, patch development and adherence to safety guardrails meant to block
Products
Brotlivulnerabilities and was able to develop and validate a critical remote code execution vulnerability in Brotli, a popular open-source data compression tool.GPT-6 CyberGPT-6 Cyber will first be available to a limited number of customers in OpenAI’s Daybreak Red program, which uses more advanced or nonpublic cybersecurity models for offensive red teaming.Linux kernelhas found at least 81 “novel” vulnerabilities in the Firefox web browser, developed exploits for 948 Linux kernel vulnerabilities and was able to develop and validate a critical remote code execution vulnerability inMozilla Firefoxmodel’s capabilities, OpenAI officials said GPT-6 has found at least 81 “novel” vulnerabilities in the Firefox web browser, developed exploits for 948 Linux kernel vulnerabilities and was able to develop and validate