In the published NexBench data, Grok 4.6 scored 114 in one recorded run, the highest single-run score. GPT-5.6 Sol's best of five runs scored 61. Grok found 44 validated issues in its run; Sol found 87 across all five. The models ran inside MindFort's agent harness.
Grok 4.6 is xAI's flagship model for long-running agentic work, released on August 12, 2026 at $2 per million input tokens and $6 per million output tokens. On NexBench it posted the highest score we have recorded. This post pairs that result with xAI's own model card, then looks at cost per run and what the safeguards still allow.
Is Grok 4.6 good at cybersecurity?
Yes. In its model card , xAI reports 79.7% on CyberGym , a benchmark of 1,507 real vulnerabilities that the model has to reproduce with a working proof of concept. That puts it just under GPT-5.6 Sol at 83.6% and ahead of Claude Opus 4.8 at 78.1%. On CVE-Bench, where an agent has to exploit real web-application CVEs in a sandbox, it scored 39.8%, up from Grok 4.5's 35.2%.
xAI says the gains are concentrated in finding and fixing bugs rather than in end-to-end attacks. Third-party evaluators given an unrestricted build confirmed its internal numbers.
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 4.6 the best model for AI pentesting?
In the current NexBench data, it has the highest single-run score: 114. This comes from one recorded Grok 4.6 run; GPT-5.6 Sol's highest score across five recorded runs was 61. Grok 4.6 submitted 45 findings and 44 survived validation, with 26 rated high severity, 18 medium, and none low. Sol had 87 accepted findings across all five runs, including 32 in its best-scoring run.
Top NexBench scores
Highest validator-accepted score each model reached at any reasoning effort, top six of 17 shown. Grok 4.6, highlighted, leads the board at 114, well clear of Grok 4.7 and GPT-5.6 Sol.
We will add runs at other reasoning efforts and update the live NexBench standings as they land.
How much better is Grok 4.6 than Grok 4.5 for pentesting?
In the recorded runs, Grok 4.6 scored 114 in its one run; Grok 4.5's best of three scored 52. Those runs found 44 and 20 validated issues, respectively. The Grok 4.6 run cost 39% less and finished 61% faster.
| NexBench result | Grok 4.5 | Grok 4.6 | Change |
|---|---|---|---|
| Score | 52 | 114 | +119% |
| Validated findings | 20 | 44 | +120% |
| High-severity findings | 14 | 26 | +86% |
| Tokens used | 26.1M | 25.4M | -3% |
| Cost per run | $62.84 | $38.63 | -39% |
| Run time | 5.04 hrs | 1.95 hrs | -61% |
xAI's own numbers show a smaller step. Its CVE-Bench score rose from 35.2% to 39.8% and CyberGym moved from 79.0% to 79.7%, which suggests the jump we measured comes mostly from how well the model runs a long engagement rather than from raw exploit skill.
How much does Grok 4.6 cost to run as a pentesting agent?
Less than any other model near the top of the board. Its NexBench run cost $38.63 at standard token prices, against $531.06 for GPT-5.6 Sol's best run, so it scored higher for about 7% of the spend. At 2.95 severity-weighted points per dollar it is the most cost-efficient model we have tested, and its $0.88 per validated finding trails only GLM-5.3 Flash.
Cost vs. token efficiency
Severity-weighted points per dollar (x) against points per million tokens (y), one point per model. Upper-right is best. Grok 4.6, highlighted, is furthest right of any model on cost efficiency.
Will Grok 4.6 refuse offensive security tasks?
Rarely, if the work is legitimate. On HackerBench v0.2, xAI's internal CTF-style suite that mixes benign tasks with dual-use and harmful requests, Grok 4.6 refused 0.0% of benign tasks and complied with 6.9% of dual-use ones, per the model card . xAI says its safeguards are tuned to keep work like vulnerability patching running.
The other frontier labs draw a tighter line. The public GPT-6 Astra refuses to write proof-of-concept exploits, and Claude Opus 5 blocks penetration testing outright.
Is Grok 4.6 better than Grok 4.7 for pentesting?
In the one recorded run for each model, Grok 4.6 scored 114 and Grok 4.7 scored 70. xAI's own Grok 4.7 model card also reports a lower CVE-Bench score, from 39.8% to 37.7%. We break down the results in our Grok 4.7 analysis.
Can you build a pentesting agent on Grok 4.6?
Yes, and its loose refusal boundary makes it one of the easier frontier models to build on. But its NexBench score came from driving our harness, not from the bare model, and xAI ran its own cyber evals the same way inside its Grok Build harness. Without that scaffolding a raw model does little against a live target: 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?
For most teams, buying wins, because the model is the easy part. Grok 4.6 makes strong offensive reasoning cheap, but a good pentesting agent harness has to scope targets, coordinate subagents across hours-long engagements, reproduce every exploit before reporting it, and stay inside the systems you authorized. That harness is why the same model can score very differently from one setup to the next.
MindFort makes that part easy. You deploy our offensive security swarms 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 4.6 good at cybersecurity?
Yes. In one recorded NexBench run, it scored 114, the highest single-run score among 17 models, with 44 validated findings and none rated low severity. xAI reports 79.7% on CyberGym and 39.8% on CVE-Bench, and says its gains are strongest at finding and fixing bugs.
How much better is Grok 4.6 than Grok 4.5 for pentesting?
In the recorded runs, Grok 4.6 scored 114 in one run. Grok 4.5's best of three scored 52. Those runs found 44 and 20 validated issues, respectively. The Grok 4.6 run cost 39% less and finished 61% faster.
Is Grok 4.6 cheaper than GPT-5.6 Sol for pentesting agents?
Much cheaper. Its NexBench run cost $38.63 against $531.06 for GPT-5.6 Sol's best run, and Grok 4.6 scored higher. At 2.95 severity-weighted points per dollar, it is the most cost-efficient model on the board.
Does Grok 4.6 refuse pentesting and red-team tasks?
Rarely, if the task is legitimate. On xAI's HackerBench v0.2, Grok 4.6 refused 0.0% of benign security tasks and complied with 6.9% of dual-use requests. That is a looser boundary than GPT-6 Astra or Claude Opus 5 draw.
Can I build a pentesting agent with Grok 4.6?
Yes, and it refuses little legitimate security work. But its NexBench score came from driving MindFort's harness, which handles the browser, the sandbox, and scoping. You would need to build that, plus a way to prove each finding before reporting it.
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.