If you've found a security issue in saklas, please report it privately rather than filing a public issue.

• Email: mx@a9l.im
• GitHub: use private security advisories

Please include a description, steps to reproduce, and the version you are on. I'll respond within a few days and aim to have a fix as soon as possible.

Only the latest minor version on PyPI receives security fixes. If you're on an older version, the fix is to upgrade.

The HTTP server (saklas serve) is meant to be used on a trusted network, like a local machine or a lab VPN. It's not meant to be used on the public internet. Please use caution.

What it does:

• Optional bearer auth via --api-key or $SAKLAS_API_KEY. If unset, the server is open.
• Per-request serialization through a single asyncio.Lock, so one slow generation doesn't interleave with another.
• Request validation via pydantic for all sampling parameters.

What it does not do:

• Rate limit
• User quotas or isolation
• Protection against adversarial inputs (logit_bias, stop, max_tokens designed to slow generation)
• TLS (please run it behind a reverse proxy if you need HTTPS)
• Sandboxing for the loaded model

If you need to expose saklas to callers you don't trust, please use a reverse proxy.

Saklas loads HuggingFace checkpoints via transformers, which executes code from the checkpoint repo in some cases. Saklas does not set trust_remote_code=True by default, but if you pass a model that requires it, please be aware that you are executing arbitrary code from that repo, so only load models from publishers you trust.

Steering vector packs pulled from HuggingFace (saklas pack install <owner>/<name>) are verified against the files sha256 map in pack.json, so on-disk tampering after download is detected. Please only install packs from publishers you trust.