Key Takeaways:
OpenAI is handing vetted cyber defenders a model that answers 95 percent of advanced hacking requests, days after delaying its Astra model over critical security risks.
OpenAI's GPT-5.6-Cyber responded to 95 percent of advanced cybersecurity requests, giving vetted defenders a tool that bypasses the guardrails limiting its flagship Sol model, which answered just 1.5 percent.
"It's definitely late," Jeffrey Ladish, executive director of Palisade Research, said of OpenAI's decision to pause Astra's development after the unreleased model neared the critical cyber capability threshold.
The new model reached only the "High" tier under OpenAI's Preparedness Framework, below Astra's critical rating. Daybreak now splits into two tiers: Blue, offering GPT-5.6 Sol without system-level cyber guardrails, and Red, granting GPT-5.6-Cyber access for exploit validation and vulnerability research. Partners Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks can embed the models into security products and managed services.
The release comes as OpenAI investigates how its own agents broke into Hugging Face, and as Microsoft's MAI-Cyber-1-Flash scored 96 percent on the CyberGym benchmark, 12 points ahead of Anthropic's Mythos.
The two-tier structure reflects the tension OpenAI faces: giving defenders real capability without leaking it to attackers. GPT-5.6-Cyber answered 95 percent of requests tied to exploit-chain development, authentication bypass and privilege escalation, while GPT-5.6-Sol responded to 1.5 percent and the Daybreak Blue version to 2 percent. The gap shows how much of the frontier model's offensive ability sits behind its safety layer.
The launch lands days after OpenAI said it was delaying Astra, its forthcoming model, after safety testing showed it could not rule out the system reaching "critical" cybersecurity capability, the highest tier under its Preparedness Framework. No prior OpenAI model has come that close. Astra is now moving into isolated testing environments with restricted network access, encrypted model weights and sandboxed execution, with government agencies and outside safety groups to test it independently before any release.
The competitive field is moving fast. Microsoft's MAI-Cyber-1-Flash, its first cybersecurity-specific model, scored 96 percent on the CyberGym benchmark when paired with GPT-5.4, 12 points ahead of Anthropic's Mythos. Anthropic released a restricted version of Mythos in June with guardrails limiting its cybersecurity and biological research capabilities, keeping the unrestricted version for a small number of vetted organizations.
For investors, the question is whether defense-focused AI becomes a durable revenue line. OpenAI's expansion of Daybreak lets security vendors embed its models into products, a direct route to commercializing cyber capability. Microsoft's roughly 27 percent stake in OpenAI is valued at about $135 billion, giving it a direct stake in how the frontier labs monetize these tools. CrowdStrike, Cisco and Palo Alto Networks, which can now resell OpenAI's cyber models, stand to gain a differentiated offering in a crowded endpoint and vulnerability market.
This article is for informational purposes only and does not constitute investment advice.
