Defense

Runtime protection: block jailbreaks and prompt injection in real time. 0DIN Defense is a fast, self-hostable classifier that scores every prompt for attack risk before it reaches your LLM, so your application can block, log, or flag it inline. The model behind it is SusFactor, 0DIN's fine-tuned prompt-suspicion classifier.

What is SusFactor?

SusFactor is a fine-tuned classifier that takes a prompt and returns a continuous suspicion score:

• score: a probability in [0, 1] that the prompt is a jailbreak or prompt-injection attempt
• is_suspicious: a boolean that's true when score >= 0.5

Detection runs entirely in-process by default, via a local PyTorch/transformers-based encoder and classification head. No prompts leave your environment, no network call required. SusFactor is trained on 0DIN's own bug-bounty threat feed of real jailbreaks disclosed by the security-researcher community, and is retrained as new attack techniques emerge.

Key Benefits

Benefit	Description
Catch attacks early	Continuous 0–1 score per prompt, so you can gate, log, or reject at the application boundary
Keep prompts in-house	Local PyTorch/transformers inference by default; nothing sent to 0DIN
Trained on real attacks	Primarily trained on real jailbreaks from 0DIN's bug-bounty threat feed, refined with WildJailbreak for a more robust perspective
Tunable per path	One score, many thresholds: strict for critical paths, looser for customer-facing flows
Lightweight	560M parameters, runs on CPU, commodity hardware
Defense-in-depth	A layer to run alongside your existing guardrails, logging, and monitoring, not a sole control

Benchmark Highlights

In benchmarking against JailbreakBench, SusFactor blocks the large majority of attacks while keeping benign false positives low. See our blog for summary benchmark numbers, and Performance for more detail.

How It Works

1. Deploy: call SusFactor via the REST API, self-hosted or through the hosted early-access beta
2. Score: score each incoming prompt 0–1 for jailbreak/injection risk, locally
3. Decide: compare the score to your per-path threshold
4. Act: gate, log, or reject the request before it reaches your LLM

Deploy Options

• Paid, self-hosted: run SusFactor yourself over REST
• Early-access beta: no infrastructure to run yourself

Early access

0DIN Defense's hosted API is currently an early-access beta. Endpoints, hostnames, and rate limits may change before general availability, and prompts sent to the hosted beta should not include production or sensitive data. Sign up for the early-access beta waitlist at 0din.ai/susfactor-trial.

Compliance Mapping

Attack categorization mapped to compliance frameworks (such as the OWASP LLM Top 10 and MITRE ATLAS) is on the roadmap. Today's API returns a suspicion score without a category breakdown; see Get Started below for what's available now.

Limitations

• Single-turn text only - SusFactor scores one prompt string per request; it doesn't evaluate multi-turn conversation history, system prompts, tool output, or RAG-retrieved content.
• Chunked scoring: prompts are split into chunks and each is scored independently; there's no length beyond which content stops being scored, but chunk boundaries mean very long prompts are evaluated piecewise rather than as one block. See API for details.
• No attack categorization - the score is a single suspicion probability, not a breakdown by attack type (see Performance).

Get Started

• API: REST endpoint, authentication, and request/response examples
• Performance: latency and benchmark numbers
• Guardrail Calibration: tuning SusFactor's threshold to your own traffic