NUWA DEVELOPER DISCLOSURE · xAI
Grok 4.7 Model Card
Developer documentation covering Grok 4.7.
Published 2026-09-21 · Updated 2026-09-21
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
CyberGym
Validated reproduction across a corpus of historical vulnerabilities.
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.
CVE-Bench
Measures agents’ exploitation of real web-application vulnerabilities in runnable environments with explicit success checks.
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.
WMDP
Multiple-choice knowledge measurements.
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.
LAB-Bench
Evaluates a range of biology research skills with structured questions; text-only or missing-input subsets remain separate protocols.
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.
StrongREJECT
A graded response evaluator for jailbreak effectiveness.
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 |
|---|---|---|
| 6 Cyber capabilities and safeguards | PDF pp.17–19 (printed 16–18) | Linked protocol Capability probes use unrestricted settings; safeguards/refusals are evaluated separately. Peer-card values are imported comparisons, not necessarily matched reruns. |
| 7 Biological and chemical capabilities and safeguards | PDF pp.20–23 (printed 19–22) | Linked protocol Knowledge/capability and refusal are explicitly different signals; only protocol metadata and bounded high-level conclusions retained. |
| 8 Jailbreaks and robustness | PDF p.23 (printed 22) | Linked protocol Production safeguards are stress-tested under single-turn, multiturn and longer attacks; residual failures motivate monitoring. |
| 11 Behaviors | PDF pp.26–27 (printed 25–26) | Contextual evidence MASK-Rectified changes grading relative to MASK, so its honesty result cannot be pooled with ordinary MASK. |
References
- [1] Grok 4.7 Model Card. 2026-09-21