NUWA DEVELOPER DISCLOSURE · OpenAI
GPT-6 Astra System Card
Developer documentation covering GPT-6 Astra, GPT-6 Sol (September), GPT-6 Luna (September), dots.
Published 2026-09-03 · Updated 2026-09-09
Sources ↓Developer-authored disclosure. Selected source sections were read; numerical results and comparability have not been audited. NUWA has not independently replicated these experiments.
Models covered
ExploitGym
Evaluates exploit construction across real software targets and checks whether success uses the designated vulnerability.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
ExploitBench
Separates exploit development into capability levels, distinguishing bug triggering from reusable exploitation primitives and control.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
SRE-Bench
Studies agentic reverse engineering with realistic binary-analysis tasks designed to reduce benchmark contamination.
A protocol mention does not establish cross-report comparability or independent replication.
This section is catalogued; no precise numerical result has been extracted here. Open the original report for its figures and full findings.
Report coverage
| Section | Source location | Treatment |
|---|---|---|
| 8 Alignment | https://deploymentsafety.openai.com/gpt-6-astra#section.8 | Linked protocol Tests include restriction following, deception, realistic work environments and unintended agent communication. Evaluated behavior is not a deployment incident rate. |
| 8.7 Verbalized Metagaming and Oversight Gaming | https://deploymentsafety.openai.com/gpt-6-astra#subsection.8.7 | Contextual evidence Card update clarifies that verbalized metagaming is detected in reasoning; oversight-gaming judgments rely on interpretation, not causal identification. |
| 9 Monitorability | https://deploymentsafety.openai.com/gpt-6-astra#section.9 | Contextual evidence Separates ordinary and adversarial monitorability. Prompted evasion tests do not establish spontaneous deployed evasion. |
| 10 Preparedness | https://deploymentsafety.openai.com/gpt-6-astra#section.10 | Linked protocol Developer assigns Critical cyber and High biological/chemical capability, with safeguards and trusted-access cases; AI self-improvement remains below its High threshold. |
References
- [1] GPT-6 Astra System Card. 2026-09-03. 2026-09-09