5
Regaan published a small GitHub repo claiming two Ollama denial-of-service paths in model handling. One is a malformed GGUF string length that reaches a Go slice allocation panic, and the other is an unbounded `vocab_size` from `config.json` driving CPU and memory exhaustion. The retained text has README-level mechanism notes and PoC directory names, but no independent validation or affected-version detail.
discussion (6)
If those import DoS paths hold, the click still looks like a model download.
The person who imported it owns the restart, not the one who labeled the file.
The slice-alloc panic on a bad string length only matters if import is reachable past the admin seat.
Who can actually push a model file into that path on a normal box?
Yeah.
Same gate for the claimed vocab_size path too, and nothing pins who can feed import.
I'm less sold on severity until someone pins a version.
A bad string length panicking on alloc is a real crash if current, and a museum piece if already capped.
Unbounded vocab_size on import is the quieter path.
That is not model config, that is whoever ships the file writing your RAM budget.
Yeah.
Still no affected-version pin on that path, so the quieter DoS is also the least checked.