When OpenAI launched GPT-6.2 Astra and Sol in November 2026, the headline numbers were about professional work: better documents, cleaner code, and a new high on multistep benchmarks. The security fine print mattered just as much. Astra became the first model OpenAI has released publicly that meets the company's High threshold for AI self-improvement, and both models meet the Critical threshold for cybersecurity. Those labels are not marketing. They are the triggers that decide what safeguards a model ships with, who can use it for sensitive work, and what a business should demand from any vendor before it connects a capable model to its systems.
Capability Thresholds Are a Safety Gauge, Not a Score
OpenAI's Preparedness Framework is the company's process for tracking advanced capabilities that could cause severe harm. It focuses on three tracked categories: biological and chemical capabilities, cybersecurity capabilities, and AI self-improvement. Capabilities are sorted into two thresholds. A High rating means the capability could amplify existing pathways to severe harm. A Critical rating means it could introduce unprecedented new pathways to severe harm. A system that reaches High must have safeguards that sufficiently minimize the associated risk before it is deployed; a system that reaches Critical must also have those safeguards in place during development.
The framework is explicit that these are dual-use categories. The same model that finds a vulnerability in your code can be pointed at someone else's. OpenAI's Safety Advisory Group, a cross-functional team of internal safety leaders, reviews whether the safeguards are enough and makes recommendations that go to company leadership. The practical meaning for a buyer is that a threshold is a statement about risk, not a badge of quality. A model crossing Critical is not automatically dangerous, but it is a model whose misuse potential has been formally acknowledged and whose deployment conditions are correspondingly stricter.
Why Capability Gains Force New Safeguards
The GPT-6.2 release is a clean example of the pattern. OpenAI says both models meet the Critical threshold for cybersecurity, meaning they find and exploit vulnerabilities more effectively, write more secure code, and produce better patches than their predecessors. They also meet the High threshold for biological and chemical capabilities. Astra goes further: it is the first externally released model to meet the High threshold for AI self-improvement, which measures a model's ability to accelerate AI research. OpenAI describes that rating as a significant step in a model's ability to accelerate AI research.
That is why the launch came with tighter safeguards for higher-risk cybersecurity assistance. OpenAI acknowledges the tension directly: the protections reduce misuse risk, but they also restrict some legitimate vulnerability research and security testing workflows. The company also built a system called Control Room to track research workloads and halt model output when intervention is needed, and expanded monitoring that can flag unauthorized behavior and pause workloads or conversations where supported. Capability and safeguard move together, and a vendor that advertises one without the other is telling you only half the story.
Daybreak Is the Gated Path for Defenders
OpenAI's answer to the tension is Daybreak, a cybersecurity program with two access tiers. Daybreak Blue covers defensive work, including vulnerability discovery, validation, remediation, threat modeling, and security investigation. Daybreak Red covers advanced, authorized red teaming, penetration testing, exploit validation, and controlled security research. Access is not automatic. OpenAI verifies applicants and reviews how they plan to use the models before approving them, and it requires stronger verification and separate approval for Red.
The tiers adjust refusals rather than removing safeguards entirely. Daybreak Blue reduces refusals on supported mainline models for verified defensive work. Access with reduced refusals runs through OpenAI's Daybreak program for approved defenders. OpenAI has also committed a billion dollars in subsidized Daybreak access, training, and support for frontline defenders at utilities, state and local governments, community banks, nonprofits, and open-source projects. For a security team, the lesson is that the most capable model is often gated behind an approval process, and that process is part of the control, not an obstacle to route around.
The Questions a Vendor Should Ask
Dual-use capability means a request can serve a legitimate defensive purpose and an offensive one, and the difference is often intent and authorization rather than wording. That makes procurement questions more important than ever. Ask what capability thresholds the model crosses and in which categories, and ask to see the evidence rather than a summary. Ask what safeguards ship by default, what they restrict, and how the vendor measures whether they hold. Ask who inside your organization can enable the higher-capability paths, what verification and hardware keys are required, and whether those permissions can be scoped to specific projects and people. Ask how refusals behave on your actual workflows, because a safeguard that blocks routine defensive work is a real operational cost. Ask what is logged, how long data is retained, and whether the vendor offers zero data retention for sensitive engagements. Ask how the model handles prompt injection and unauthorized instructions, and what monitoring alerts you will receive. A vendor that treats these questions as normal is a vendor you can work with.
Using a Capable Model Responsibly
The same discipline applies inside your own walls. Give the model the least access it needs to do the job, and keep it on systems you own or are explicitly authorized to test. Treat its output as a strong draft rather than a verified fact, especially for security findings, and require a human to validate any exploit or patch before it runs anywhere. Separate the high-capability paths from customer-facing and third-party traffic, and keep them on internal projects with named owners. Monitor for overreach as well as failure, because a capable agent can stay quiet about the obstacles it hit and the steps it skipped. Document what you authorized and what the model did, so an incident has a trail. Capability without that scaffolding is how a useful tool becomes a liability.
The models arriving in 2026 are powerful enough that security can no longer be an afterthought bolted on at deployment. If you want a practical framework for evaluating AI models, their tradeoffs, and the safeguards they require, the team at Crowe Endeavors publishes guides and tools at https://crowe-endeavors.com.
Sources
"AI News Brief – November 18, 2026" (Trend Match Touch (YouTube), November 18, 2026) — https://www.youtube.com/watch?v=4FUUyQuViJA
"Introducing GPT-6.2 Astra and Sol (cyber thresholds and Daybreak)" (OpenAI, November 17, 2026) — https://openai.com/index/introducing-gpt-6-2-astra-sol/
"OpenAI releases GPT-6.2 Astra and Sol models" (SCWorld, November 18, 2026) — https://www.scworld.com/brief/openai-releases-gpt-6-2-astra-and-sol-models