unplug goes live

unplug-ai.org is live. Everything lives under the UnplugAI org now: SDK, site, models, datasets.

// what's shipping

The SDK has a 3-stage detection pipeline, taint tracking across execution graphs, and a 12-stage text normalizer that handles evasion techniques most tools skip. We're training Unplug Tiny, our own lightweight classifier, on a synthetic + curated high-quality dataset that we're actively scaling. Tiny ships first, then Small, then larger variants.

The plan is to release Unplug Tiny fully open source alongside the SDK. The models and the dataset both go public.

// what's next

• Ship Unplug Tiny as an open-source model on HuggingFace
• Publish the SDK to PyPI with the classifier baked in
• Launch the hosted API as a service for teams that don't want to self-host
• Framework integrations, starting with Agno
• Scale the dataset and train larger model variants

// why this matters

AI agents now execute code, query databases, and move money. None of the major frameworks have built-in security. Four companies in this space got acquired for a combined $1.3 billion in the last year. The independent open-source position is wide open.

Site: unplug-ai.org. Code: github.com/UnplugAI.