DeepSeek V4.1 Flash showed a result of 84 points in the hacker benchmark

The DeepSeek V4.1 Flash model demonstrated high efficiency in automated vulnerability search, scoring 84 points in a specialized hacking benchmark. The testing process took 4 hours, during which the neural network successfully executed code on target systems.

As part of the testing, the model achieved a result of 11/11 in attacks on platforms such as Grafana and Jenkins. The total cost of performing all operations was $4.65. Researchers noted that five successful scenarios were implemented through undocumented routes of the test environment, rather than through target vulnerabilities, which indicates the model's features in non-standard conditions.

The testing results of DeepSeek V4.1 Flash highlight the current capabilities of specialized LLMs in cybersecurity tasks. The use of such tools allows for the automation of attack vector search with minimal computational resource costs.

This benchmark is part of the efforts to evaluate autonomous agents in real software exploitation scenarios. The obtained data help developers better understand the boundaries of model applicability when working with critical infrastructure and complex software systems.