Tesla’s AI stack powers Full Self-Driving, Optimus humanoid robots, and real-world autonomy with neural networks, custom hardware, and Dojo-scale training.
31M visits/mo+5.5%#8 Coding & Development
unplug
LLM defense platform that blocks prompt injection and data leakage in real time
FreeStarting price
About unplug
unplug is a defense platform designed to protect large language model (LLM) applications from adversarial attacks such as prompt injection, destructive actions, and data exfiltration. It operates as a three-stage pipeline—normalize, scan, and decide—processing inputs in under 20 milliseconds to identify and block malicious content before it reaches the agent. The system supports both local SDK deployment for offline, CPU-only scanning and a planned hosted API for server-side integration. It provides span-level redaction, evidence-based findings with risk scores, and source-aware scanning across user inputs, retrieved content, tool outputs, and system prompts. The platform categorizes threats into six scanner types, including injection, destructive actions, data leakage, harmful outputs, financial risks, and secret exposure. Benchmarks show high recall and precision rates, particularly for indirect injection attacks hidden in retrieved context, with false positive rates below 1%.
Key features
- Local SDK with pip install support
- Span-level redaction of malicious content
- Evidence-based findings with risk scores
- Source-aware scanning for multiple input types
- Six threat categories: injection, destructive, leakage, harmful, financial, secrets
- Global instance initialization with Guard.init()
- Offline and CPU-only operation in local mode
- Planned server mode with tier classifiers
Use cases
- Blocking prompt injection attacks in LLM applications
- Preventing data exfiltration of API keys or PII
- Mitigating destructive commands like DROP TABLE or rm -rf
Pros
- Three-stage pipeline with sub-20ms latency
- Span-level redaction and evidence-based findings
- Supports local SDK and planned hosted API
- Six scanner categories for comprehensive threat coverage
- Source-aware scanning for user, retrieved, tool, and system inputs
Cons
- Hosted API not yet live (founding waitlist only)
- Python SDK only (JavaScript, Go, Rust coming soon)
- Single-turn detection only (multi-turn trajectory detection measured separately)
- No free tier or open-source release
Frequently asked questions about unplug
What does unplug do?
unplug is an LLM defense platform that scans untrusted text for prompt injection, destructive actions, and data leakage. It provides evidence-based findings and span-level redaction rather than binary blocking.
Who is unplug designed for?
unplug is designed for developers and organizations deploying large language model applications that require protection from adversarial attacks such as prompt injection and data exfiltration.
How can I integrate unplug into my workflow?
unplug can be integrated via a free Python SDK available on PyPI, with installation options including basic and ML-enhanced versions. A hosted API is planned for server-side integration.
Does unplug support agent-based systems?
Yes, unplug offers an MCP server for agent-based systems, available via uvx unplug-mcp. This requires the unplug-ai SDK version 0.5.2 or higher.
What types of threats does unplug detect?
unplug detects prompt injection, destructive actions (such as SQL or shell commands), data leakage (including API keys and PII), harmful outputs, financial risks, and secret exposure.
Can unplug be used offline?
Yes, unplug supports local SDK deployment for offline, CPU-only scanning, making it suitable for environments with limited connectivity or strict data privacy requirements.
