DeepSeek V4.1-Flash-1017 is the best pentesting model we have tested. DeepSeek's October 17 release scored 121 on NexBench, first of 19 models and 6% above Grok 4.6. Its best run cost $25.50. It was also the cheapest model per validated finding, at $0.60. Its open weights put the same skill in attackers' hands.
DeepSeek V4.1 Flash is DeepSeek's open-weight model, released on September 10, 2026 under the MIT license. DeepSeek shipped its latest release, V4.1-Flash-1017 , on October 17. We ran that release through NexBench at four effort levels, and it took first place. This post covers what it scored and what it costs, plus the risks of its open weights.
Is DeepSeek V4.1-Flash-1017 good at cybersecurity?
DeepSeek's own numbers for the earlier release say yes. The September release scored 88.1% on CyberGym, ahead of GPT-5.6 Sol and GLM-5.3 at 84.5%, per its model card . It scored 15.3% on ExploitGym, though, well behind Sol's 33.7%.
The October release has no published cyber scores. DeepSeek says it finishes tasks faster with shorter outputs, per its release notes . On Terminal-Bench 4.0 it rose from 31.2% to 42.3%, per its model card .
How do you benchmark an LLM for pentesting?
We use NexBench, MindFort's internal benchmark for AI models doing offensive security work. Each model runs inside a portable version of our production agent harness. It then runs an authorized pentest of a realistic web app. That tests its decision-making, execution, and reasoning in a real-world environment.
An independent GPT-5.5 validator reproduces every finding before it counts. High and critical findings score 3 points, medium findings score 2, and low findings score 1. Each run stops at a five-hour cap, and the full method is in our NexBench write-up.
Is DeepSeek V4.1-Flash-1017 good for AI pentesting?
Yes. Its best NexBench run scored 121 at high effort, the top score among 19 models. It edged past Grok 4.6's 114 and nearly doubled GPT-5.6 Sol's best of 61.
Top NexBench scores
Highest validator-accepted score each model reached at any reasoning effort, top six of 19 shown. DeepSeek V4.1-Flash-1017, highlighted, ranks first ahead of Grok 4.6.
Across four runs, 135 of its 140 submissions passed validation, or 96%. Of those, 89 were high severity, the most of any model on the board. Even with thinking turned off, it scored 52 in 49 minutes, above Opus 4.8's best run. The live NexBench standings will update as we add runs.
Is DeepSeek V4.1-Flash-1017 better than Grok 4.6 for pentesting?
Narrowly on score, and clearly on cost. Its best run scored 6% higher than Grok 4.6's and cost 34% less. It took much longer, though, and ran to the five-hour cap.
| NexBench best run | Grok 4.6 | DeepSeek V4.1-Flash-1017 | Change |
|---|---|---|---|
| Score | 114 | 121 | +6% |
| Validated findings | 44 | 48 | +9% |
| High-severity findings | 26 | 29 | +12% |
| Rejected findings | 1 | 0 | -100% |
| Cost per run | $38.63 | $25.50 | -34% |
| Run time | 1.95 hrs | 5 hrs | +156% |
We have not run an earlier DeepSeek model on NexBench, so there is no same-family comparison yet. The closest is the last open-weight leader, Kimi K3, whose best run scored 42. V4.1-Flash-1017 beat it by 188% for about $6 more per run.
Score vs. cost per run
Every model-effort run by standard cost against score. DeepSeek V4.1-Flash-1017's four runs, highlighted with Grok 4.6, all cost under $28.
Which DeepSeek V4.1-Flash-1017 effort level is best for pentesting?
High, which is the API default. It scored 121 and found 48 validated bugs with zero rejections. Max scored 120 and finished in 3.4 hours, so it is the better pick when time matters. Max also found the most high-severity bugs, 32 to high's 29.
| Effort | Score | Validated findings | High severity | Run time | Cost |
|---|---|---|---|---|---|
| Thinking off | 52 | 22 | 12 | 49 min | $13.30 |
| Low | 58 | 21 | 16 | 53 min | $14.54 |
| High | 121 | 48 | 29 | 5 hrs | $25.50 |
| Max | 120 | 44 | 32 | 3.42 hrs | $27.47 |
Turning thinking up mattered more than anything else. Going from low to high more than doubled the score, from 58 to 121. DeepSeek maps medium and xhigh requests to high, per its API docs , so these four levels are all it offers.
How much does DeepSeek V4.1-Flash-1017 cost to run as a pentesting agent?
Less than any model we have tested. All four NexBench runs together cost $80.81, or $0.60 per validated finding. Claude Haiku 5.5 came second at $0.70.
The API charges $0.30 per million input tokens and $1.20 per million output tokens at peak hours, per DeepSeek . Off-peak rates are half that. The open weights also let you run it on your own hardware.
Should you build your own pentesting agent or buy one?
Buy, unless you want to own the hard part. V4.1-Flash-1017 shows how much a cheap open model can do inside a good harness. Building that harness is hard. It has to scope targets, coordinate subagents for hours, and reproduce every exploit inside authorized systems.
MindFort makes that part easy. Our pentesting swarms run on MF-1, our custom LLM built for offensive security reasoning, inside the same harness NexBench is built on. Every proven finding comes back as a merge-ready GitHub PR. Multiple Fortune 500 companies and top startups already use them, and you can book a demo to see them on your stack.
FAQ
Is DeepSeek V4.1-Flash-1017 good at pentesting?
Yes. DeepSeek's October 17 release scored 121 on NexBench, the top score among 19 models. It beat Grok 4.6's 114 and nearly doubled GPT-5.6 Sol's 61. Of its 140 submitted findings, 135 passed validation.
Is DeepSeek V4.1-Flash-1017 better than Grok 4.6 for pentesting?
On NexBench, yes. Its best run scored 121 to Grok 4.6's 114 and cost 34% less. It was slower, though. That run went to the five-hour cap, while Grok 4.6 finished in under two hours.
How much does DeepSeek V4.1-Flash-1017 cost for pentesting?
Its API lists at $0.30 per million input tokens and $1.20 per million output tokens at peak hours. Off-peak rates are half that. On NexBench it cost $0.60 per validated finding, the lowest of any model.
Are there uncensored versions of DeepSeek V4.1-Flash-1017?
Yes. Community builds on Hugging Face strip the safety refusals from both the 1017 release and the earlier V4.1 Flash. One removes the cybersecurity offense and defense rails. The MIT license lets anyone download and run them.
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.