Enclave Says 5 of 11 DeepSeek V4.1 Flash Wins Skip Intended Bugs
Summary
Enclave says DeepSeek V4.1 Flash verified 11 of 11 vulnerable targets on its AI Hacking Race benchmark in September, but five wins skipped the bugs the challenges were built around. Enclave closed the shortcut routes and versioned the benchmark. A rerun on the new version verified 4 of 11.
Key Points
- Enclave co-founder Yanir Tsarimi reports DeepSeek issued 2,349 Bash commands across 2 hours 38 minutes in the September run.
- Three Grafana runs need only 52, 64 and 90 seconds to reach the flag, skipping the intended file path bug.
- Enclave reruns the model on DeepInfra's FP8 endpoint because Novita's does not host the newer 1017 snapshot.