OpenAI Gives Security Teams Access to a Model That Rejects Fewer Risky Cyber Tasks

OpenAI introduced GPT-5.6-Cyber and at the same time divided its Daybreak security program into two operating modes. Daybreak Blue gives approved defenders access to general models, including GPT-5.6 Sol, with adjusted safety restrictions. Daybreak Red goes further and allows the use of specialized models for vulnerability research, exploit validation, and other tasks that have both legitimate and offensive applications. The core change is precisely the reduction in refusal behavior. A normally deployed model may reject an instruction that looks like an attempt to bypass authentication or escalate privileges even when it is being issued by a security team during an approved test. According to OpenAI, GPT-5.6-Cyber is trained to complete such tasks more often within a controlled environment. OpenAI's internal Advanced Cybersecurity Completion Rate test includes scenarios such as developing exploit chains, bypassing authentication, and escalating privileges. GPT-5.6-Cyber completed 95 percent of requests, according to the company. Standard GPT-5.6 Sol completed 1.5 percent and Sol in Daybreak Blue mode completed two percent. The previous GPT-5.5-Cyber reached 57.3 percent. The difference is therefore not cosmetic; the specialized version is intentionally far less restrictive. This creates an important risk-governance problem. The ability to find a flaw and produce a working proof of exploitation is useful to defenders, but the same procedure can be used by an attacker. Safety therefore cannot be handled only through a filter inside the model. OpenAI is moving part of the protection to the level of user identity, access approval, and a controlled operating environment. For enterprise security teams, this could shorten the time between discovering a weakness and verifying its real severity. AI can review source code, look for combinations of flaws, assemble test scenarios, and prepare remediation proposals. The benefit, however, depends on the processes around the model: request logging, separation of test and production environments, approval of dangerous actions, and clear human accountability. The area is also important from a regulatory perspective because companies will increasingly need to demonstrate not only which model they use but also who has access to its most dangerous capabilities. The difference between ordinary and privileged modes may resemble the way organizations manage administrator privileges in information systems. OpenAI describes its goal as giving advanced capabilities to defenders before attackers can use them at scale. The company also warns that “defenders have a narrowing window to prepare”. Selective access to more dangerous capabilities is likely to become one of the key experiments in governing increasingly powerful AI systems. Key terms Exploit: A program or procedure that takes advantage of a specific system vulnerability. Zero-day vulnerability: A security flaw for which no fix is available when the flaw is discovered. Dual use: A technology or capability that can serve both legitimate and harmful purposes.

OpenAI Gives Security Teams Access to a Model That Rejects Fewer Risky Cyber Tasks

The full article is available to users registered at Hard Skills