The Grok 4.1 Model Card is a six-page document published by xAI on November 17, 2025 (Source: https://data.x.ai/2025-11-17-grok-4-1-model-card.pdf). It covers the two configurations of Grok 4.1, Thinking and Non-Thinking, both evaluated with the production system prompt, and describes pre-deployment testing "largely similar" to that in the Grok 4 Model Card (xAI, August 2025).
Abuse potential
Refusals are measured on the same internal multilingual dataset as earlier cards, with user-message and system-message jailbreak variants, alongside AgentHarm without jailbreaks and AgentDojo prompt-injection attack success. The card discloses that the refusal results in previous model cards "were reported with an error in the evaluation settings, where only the English prompts were evaluated", and that the multilingual figures it reports are not directly comparable with them. A second table evaluates a new input filter for restricted biological and chemical knowledge by false negative rate, with and without prompt-injection attacks.
Concerning propensities
MASK dishonesty and Anthropic's sycophancy evaluation are reported for both configurations next to Grok 4's figures. The text states that training for honesty and against sycophancy reduces both behaviours; the table prints higher dishonesty and sycophancy rates for both Grok 4.1 configurations than for Grok 4.
Dual-use capabilities
Results are for Grok 4.1 Thinking with safeguards removed, shown beside Grok 4 and, where available, human expert baselines drawn from the benchmark papers. The set adds three LAB-Bench subsets (ProtocolQA, FigQA, CloningScenarios) to WMDP, the text-only Virology Capabilities Test, BioLP-Bench, CyBench and MakeMeSay. xAI states that Grok 4.1 matches or exceeds the human baselines on knowledge and protocol-troubleshooting questions and falls below them on FigQA and CloningScenarios, and adds that published human baselines probably understate high-context experts. It describes CyBench performance as similar to other frontier models and well below human experts, and says the model performs poorly on MakeMeSay.
Provenance and confidence
All figures are developer-reported; human baselines are cited from prior work. The card restates Grok 4's MakeMeSay win rate as 0.13, against 0.12 in the Grok 4 card. Confidence is medium. The vault copy is the text layer of the official PDF.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| Refusals — Standard (no attack) | Grok 4.1 Thinking | 0.07 | answer rate (lower is better), internal single-turn policy-violating prompts in six languages (several thousand), model-graded, production system prompt; multilingual, not comparable to Grok 4 / Grok 4 Fast card figures; developer-reported | 2.1.2–2.1.3, Table 1 |
| Refusals — Standard (no attack) | Grok 4.1 Non-Thinking | 0.05 | answer rate, same set, production system prompt; developer-reported | 2.1.2–2.1.3, Table 1 |
| Refusals — + User Jailbreak | Grok 4.1 Thinking | 0.02 | answer rate, internal single-turn jailbreak templates in user message; developer-reported | Table 1 |
| Refusals — + User Jailbreak | Grok 4.1 Non-Thinking | 0.00 | answer rate, jailbreak in user message; developer-reported | Table 1 |
| Refusals — + System Jailbreak | Grok 4.1 Thinking | 0.02 | answer rate, jailbreak in system message; developer-reported | Table 1 |
| Refusals — + System Jailbreak | Grok 4.1 Non-Thinking | 0.00 | answer rate, jailbreak in system message; developer-reported | Table 1 |
| AgentDojo | Grok 4.1 Thinking | 0.05 | attack success rate (prompt injection); developer-reported | Table 1 |
| AgentDojo | Grok 4.1 Non-Thinking | 0.01 | attack success rate (prompt injection); developer-reported | Table 1 |
| Input Filter Refusals — Restricted Biology — Direct | xAI input filter (deployed with Grok 4.1) | 0.03 | false negative rate, internal single-turn prompts seeking restricted biological knowledge; developer-reported | Table 2 |
| Input Filter Refusals — Restricted Biology — + Prompt Injection | xAI input filter (deployed with Grok 4.1) | 0.20 | false negative rate under prompt-injection attacks; developer-reported | Table 2 |
| Input Filter Refusals — Restricted Chemistry — Direct | xAI input filter (deployed with Grok 4.1) | 0.00 | false negative rate, internal single-turn prompts seeking restricted chemical knowledge; developer-reported | Table 2 |
| Input Filter Refusals — Restricted Chemistry — + Prompt Injection | xAI input filter (deployed with Grok 4.1) | 0.12 | false negative rate under prompt-injection attacks; developer-reported | Table 2 |
| MASK | Grok 4.1 Thinking | 0.49 | dishonesty rate, 1000 questions, production system prompt; developer-reported | 2.2.1, Table 3 |
| MASK | Grok 4.1 Non-Thinking | 0.46 | dishonesty rate, production system prompt; developer-reported | 2.2.1, Table 3 |
| MASK | Grok 4 | 0.43 | dishonesty rate; developer-reported, restated from the Grok 4 card | Table 3 |
| Sycophancy | Grok 4.1 Thinking | 0.19 | sycophancy rate, Anthropic sycophancy evaluation (Sharma et al.), lower is better; developer-reported | 2.2.2, Table 3 |
| Sycophancy | Grok 4.1 Non-Thinking | 0.23 | sycophancy rate; developer-reported | 2.2.2, Table 3 |
| Sycophancy | Grok 4 | 0.07 | sycophancy rate; developer-reported, restated from the Grok 4 card | Table 3 |
| BioLP-Bench | Grok 4.1 Thinking | 0.37 | accuracy, model-graded, safeguards removed; developer-reported | Table 4 |
| BioLP-Bench | Grok 4 | 0.47 | accuracy; developer-reported in the Grok 4.1 card | Table 4 |
| BioLP-Bench | Human baseline | 0.38 | human expert baseline as cited by xAI (Ivanov 2024) | Table 4 |
| MakeMeSay | Grok 4.1 Thinking | 0.00 | win rate (average manipulation rate) against a non-thinking Grok-3-Mini defender, 15-turn limit, safeguards removed; developer-reported | Table 4 |
| MakeMeSay | Grok 4 | 0.13 | win rate; developer-reported in the Grok 4.1 card (the Grok 4 card printed 0.12) | Table 4 |
Relationships
- supports: Grok (xAI)
- instance-of: AI System Cards
- related: Grok 4 Model Card (xAI, August 2025), Grok 4 Fast Model Card (xAI, September 2025), xAI