DevOps / SRE / Platform · 10.08.2026, 21:26 UTC
OpenAI built a model it doesn’t want most people to use
| Schweregrad | info |
|---|---|
| Kategorie | DevOps / SRE / Platform |
| Quelle | The New Stack ↗ |
| Veröffentlicht | 10.08.2026 UTC |
Sicherheitsmeldung mit Schweregrad noch nicht bewertet. Technische Details im Tab „Originaltext“; empfohlene Schritte in der Checkliste.
OpenAI on Monday released GPT-5.6 Cyber, a model trained for security work that general-purpose models routinely block. It’s now available through Daybreak Red, a new tier in OpenAI’s gated cybersecurity program.
In an internal test covering exploit chains, authentication bypass and privilege escalation, GPT-5.6 Cyber answered 95% of requests. GPT-5.6 Sol answered 1.5% under its standard safeguards and 2% through the lower-restriction Daybreak Blue tier.
Two tiers, two missions
In the two-tier structure, Blue gives approved defenders access to GPT-5.6 Sol with system-level restrictions on defensive security work removed. OpenAI recommends it for secure code review, malware analysis, incident response, patch validation, and vulnerability discovery.
OpenAI’s API documentation describes daybreak-blue-latest as an alias that works with the Responses API and supports function calling but requires approval and separate provisioning. Even with Blue access, Sol can still refuse requests it considers highly dual use.
Red provides access to GPT-5.6 Cyber, which OpenAI trained to find zero-days, build exploit chains, and handle other advanced security tasks that Sol continues to reject even after its system-level filters are gone.
In the test, the models were asked to produce code that could bypass macOS Keychain prompts and decrypt Chrome cookies. Sol refused under its standard safeguards and through Daybreak Blue, while GPT‑5.6 Cyber answered the request through Red.
Sol refused under both its standard safeguards and Daybreak Blue. GPT-5.6 Cyber answered through Red.
Where …