NUWA DEVELOPER DISCLOSURE · Google DeepMind
Gemini 4 Argon Safety & Security Report
Developer documentation covering Gemini 4 Argon.
Published Not verified
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
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.
Report coverage
| Section | Source location | Treatment |
|---|---|---|
| 3.2 Misuse risk assessment | PDF printed pp.21–53 | Linked protocol Separates inherent capability from residual risk and uses internal and independent tests; no operational biology details or numerical scores are transcribed. |
| 3.3 Misuse Safeguards | PDF printed pp.55–62 | Contextual evidence Coverage and adversarial robustness are assessed separately; safeguards are part of the deployment risk argument. |
| 3.4 Loss of Control | PDF printed pp.63–155 | Contextual evidence Holistic safety argument combines capabilities, propensities, monitoring and security assumptions; conclusions depend on those stated assumptions. |
| 3.4.5 Chain-of-Thought Transparency Assessment | PDF printed pp.112–155 | Contextual evidence Developer reports more difficult monitorability than prior models; complex covert-task results are bounded by task difficulty and monitor strength. |
| 3.5 Evaluation Awareness and Behavior Effects | PDF printed pp.156–158 | Contextual evidence Awareness can change compliance and local grader behavior. Audits found no confirmed strategic underperformance; this does not rule out unspoken awareness. |
| 6 Guarding users against harm from third parties | PDF §6 | Contextual evidence Prompt-injection resistance is a separate model/system robustness finding. |