Skip to content
Hacker News front page

DeepSeek V4.1 Flash scores 11/11 on Enclave's hacking benchmark

DeepSeek v4.1 Flash Is Now Our Best Hacking Model

Enclave tested DeepSeek V4.1 Flash against 11 vulnerable targets and 4 patched controls. The model achieved code execution on all 11 targets for $4.65 total, using 2,349 Bash commands over nearly 2h38m of active model time. An audit confirmed 6 attacks followed the planned exploit path; 5 found alternate routes in the test environment, including a Grafana compromise in 52 seconds. Enclave has since closed those extra paths and will require re-runs for updated leaderboard rankings.

Why it matters: Enclave reports DeepSeek V4.1 Flash hit code execution on all 11 vulnerable targets for $4.65, while its audit found five attack paths the scoring system did not distinguish.

Read the original ↗Export Markdown