OpenAI Expands Daybreak to Tackle Growing AI Security Threat

AlexSava via Getty ImagesOpenAI is expanding the scope of Daybreak, the cybersecurity initiative that the AI lab introduced in May as a rival to Anthropic’s Project Glasswing.The extension of Daybreak was confirmed in a blog update on the OpenAI website, which revealed that the program will now have two different access tiers -- Daybreak Blue and Daybreak Red.The change reflects a landscape in which concern is escalating following a number of cybersecurity incidents over the past few weeks, in which models from OpenAI itself, Anthropic and Meta all accessed areas online that should have been out of bounds from their testing environments.And last week, OpenAI paused testing on another model, Astra, because it had reached its “critical cybersecurity threshold.”In the post OpenAI admitted the threat from AI was growing, saying: “The cybersecurity world is rapidly changing -- threat actors will increasingly use AI to conduct cyberattacks at unprecedented speed and scale, including in fully autonomous ways. As these capabilities spread, defenders have a narrowing window to prepare.”Related:Security Concerns Cause OpenAI to Halt Work on Astra ModelOpenAI’s response is to give the new Daybreak Blue tier users access to frontier general-purpose models, including GPT‑5.6 Sol, with safeguards tweaked to allow for authorized defensive security work. Daybreak Blue is the more basic of the two new tiers and is the recommended starting point for most defenders.Those using the more advanced Daybreak Red, meanwhile, will get access to OpenAI’s purpose-trained cybersecurity models for testing and vulnerability research. They will also have access to GPT‑5.6‑Cyber, a new model built on GPT‑5.6 Sol that has been trained to deliver improved capabilities across several specialized cybersecurity tasks and is less likely to refuse high-risk tasks.According to OpenAI, making the model less likely to refuse certain cybersecurity tasks addresses “feedback from security researchers who encountered persistent refusals” from GPT‑5.5‑Cyber for requests involving advanced cybersecurity scenarios. GPT‑5.6‑Cyber completes 95.0% of requests, well up from the previous model’s 57.4%.As has been the case since Daybreak was launched, access is restricted to approved organizations and individuals performing authorized work, with OpenAI citing the potential risks of misuse.However, the vendor appears ready to expand access to Daybreak, with the blog containing a link to a page where potential users can submit an official application to sign up.About the AuthorContributing WriterGraham Hope has worked in automotive journalism in the U.K. for 26 years, including spells as editor of leading consumer news website and weekly Auto Express and respected buying guide CarBuyer.