Grok 4.7, released by xAI on September 21, 2026, is a strong security model, but it is a step back from Grok 4.6 at live web-app pentesting. It scored 70 on MindFort's NexBench, second of 17 models, against Grok 4.6's 114. Every finding it submitted held up, and 77% were high severity.
Grok 4.7 is xAI's newest flagship, released on September 21, 2026 at the same $2 per million input tokens and $6 per million output tokens as Grok 4.6. We ran it through NexBench and checked the results against xAI's model card. Both point the same way: a small gain on hard exploit chains, a slight dip on web-app exploitation, and much heavier token use.
How do you benchmark an LLM for pentesting?
We use NexBench, MindFort's internal benchmark for AI models doing offensive security work. Each model is placed inside a portable version of our production agent harness, which gives it a browser, an isolated container, and a manager that hands work to subagents. It then runs an authorized full-breadth pentest of a realistic web app, so the test measures the model's decision-making, execution, and reasoning in a real-world environment rather than on isolated puzzles.
An independent GPT-5.5 validator reproduces every finding before it counts. Scores are severity-weighted at 3 points for a high or critical finding, 2 for medium, and 1 for low, and the full method is in our NexBench write-up.
Is Grok 4.7 good for AI pentesting?
Yes. It scored 70 on NexBench, second of 17 models and ahead of GPT-5.6 Sol's best run of 61. All 26 findings Grok 4.7 submitted passed validation, and 77% of them were high severity, the highest share on the board.
Top NexBench scores
Highest validator-accepted score each model reached at any reasoning effort, top six of 17 shown. Grok 4.7, highlighted, ranks second behind Grok 4.6 and ahead of GPT-5.6 Sol.
The live NexBench standings will update as we add runs.
Is Grok 4.7 better than Grok 4.6 for pentesting?
Not in our testing. Grok 4.7 scored 70 to Grok 4.6's 114, a 39% drop, and its validated findings fell 41%, from 44 to 26. It also used 39% more tokens to get there.
| NexBench result | Grok 4.6 | Grok 4.7 | Change |
|---|---|---|---|
| Score | 114 | 70 | -39% |
| Validated findings | 44 | 26 | -41% |
| High-severity findings | 26 | 20 | -23% |
| Tokens used | 25.4M | 35.2M | +39% |
| Cost per run | $38.63 | $43.50 | +13% |
| Run time | 1.95 hrs | 1.28 hrs | -34% |
xAI's own model card shows a similar split. CyberGym ticked up from 79.7% to 80.3%, but CVE-Bench, its web-app exploitation test, fell from 39.8% to 37.7%. Where Grok 4.7 gains is depth: it scored 29% on the hard subset of CathedralBench, a third-party test of multi-exploit chains, against Grok 4.6's 25%, and in our run its high-severity share rose from 59% to 77%.
Score vs. cost per run
Every model-effort run by standard cost against score. Grok 4.7 and Grok 4.6, highlighted, both sit in the cheap, high-scoring corner, with Grok 4.6 higher on score and lower on cost.
Why does Grok 4.7 use more tokens than Grok 4.6?
It thinks for longer. xAI says Grok 4.7 spends more time on hard problems and checks its own answers more often, per Decrypt . Artificial Analysis measured about 81,000 output tokens per Intelligence Index task at xhigh effort, more than double Grok 4.6's 36,000, per Metaverse Post .
On NexBench the extra thinking added 13% to the bill. It still finished 34% faster than Grok 4.6, which fits a model that chases fewer leads and spends more on each one.
Will Grok 4.7 refuse offensive security tasks?
Some of them. It blocks more risky requests than Grok 4.6 while turning away almost no benign ones. On HackerBench v0.3, xAI's internal CTF-style suite, Grok 4.7 let through 3.3% of risky dual-use prompts, per xAI . Grok 4.6 let through 6.9% on the previous version of the suite. xAI has also started giving select security partners invite-only access to Grok 4.7's red-team capabilities for defense research.
The model card also says xAI never silently downgrades to a different model. That is a real difference from Claude Opus 5, which reroutes flagged cyber requests to Opus 4.8.
Can you build a pentesting agent on Grok 4.7?
You can, but the model is the smaller part of the job. The public version refuses the riskiest offensive work and the unrestricted red-team mode is invite-only, so your agent has to work inside those limits. You also have to build what a raw model lacks, like target scoping, exploit validation, and evidence capture. That harness decides most of the outcome: the same GPT-4 that exploited 87% of one-day vulnerabilities inside an agent harness managed 7% without one .
Should you build your own pentesting agent or buy one?
Buy, unless you want to own the harness. Grok 4.7 shows a model can even get worse at pentesting between versions, which is why we re-run every release on NexBench, as we did for Grok 4.6 and GPT-6 Astra. The harder part is the harness itself: it has to scope targets, coordinate subagents across hours-long engagements, reproduce every exploit before reporting it, and stay inside the systems you authorized.
MindFort makes that part easy. You deploy our pentesting agents against your apps, APIs, and infrastructure, where they run on MF-1, our custom LLM built for offensive security reasoning inside the same production harness NexBench is built on, and every proven finding comes back as a merge-ready GitHub PR. Multiple Fortune 500 companies and top startups already rely on them, and you can book a demo to see them run against your stack.
FAQ
Is Grok 4.7 good at cybersecurity?
Yes. It scored 70 on MindFort's NexBench, second of 17 models and ahead of GPT-5.6 Sol's best run of 61. All 26 findings it submitted passed validation, and 77% were high severity, the highest share of any model we have tested.
Is Grok 4.7 better than Grok 4.6 for pentesting?
Not in our testing. Grok 4.7 scored 70 on NexBench, 39% below Grok 4.6's 114, with 41% fewer validated findings and 39% more tokens used. xAI's own model card shows Grok 4.7 slipping on CVE-Bench, its web-app exploitation test, from 39.8% to 37.7%.
Does Grok 4.7 refuse pentesting and red-team tasks?
Rarely, if the work is legitimate. xAI rebuilt the safeguard stack, and on HackerBench v0.3 Grok 4.7 let through 3.3% of risky dual-use prompts while rarely blocking legitimate security work. Its full red-team capabilities are invite-only for select security partners.
Can I build a pentesting agent with Grok 4.7?
Yes, within limits. The public model refuses the riskiest offensive work and the unrestricted red-team mode is invite-only. You also have to build the scoping, exploit validation, and guardrails that keep an agent inside your authorized targets.
About the author
Brandon Veiseh
Co-Founder & CEO · MindFort
Founded his first startup building NLP models for network packet inspection. Led product at ProjectDiscovery, built their enterprise platform from scratch. At NetSPI, led development of AI tools for offensive security.