An AI security firewall for LLM applications, built around a six-layer detection pipeline: normalization, regex, ML ensemble, LLM-based content safety, multi-turn intent drift, and policy. It sits in front of OpenAI, Anthropic, Gemini, Azure, Bedrock, Mistral, Cohere, DeepSeek, Groq, Hugging Face, Ollama, and vLLM, judging each request before it reaches the model. Deployment comes in three modes—managed cloud (swap a base URL), hybrid self-host (Helm or docker-compose, with metadata forwarding so no prompt text leaves), and fully air-gapped (offline licence validation, egress denied). A separate Shadow AI product extends the same engine to staff-facing tools through a browser extension and macOS/Windows agent, masking PII on-device. Published benchmarks cover 2,369 samples across 8 datasets (5 independently published): precision 99.1%, recall 80.3%, over-defense 45.4%, F1 0.887. Pricing is usage-based: Free includes $5 of usage per month with no credit card, Developer is $20 of usage per month, Enterprise is custom with SSO, audit logs, and self-hosting. PromptGuard is the trading name of PG Tech Ltd, registered in England and Wales (company number 17437377).
Pros & Cons:
Pros:
Six-layer pipeline with a public, reproducible benchmark (F1 0.887, 2,369 samples, 8 datasets, confidence intervals published) rather than an unsourced detection claim; three deployment modes including fully air-gapped, which is rare for AI security tooling; the gateway works by swapping a base URL or adding one SDK line, so existing provider code does not need rewriting; Shadow AI covers staff use of ChatGPT, Claude, Gemini, Copilot and Cursor, masking PII on-device before transmission; explicit, unusually candid compliance statement—“Not certified, and we will not imply otherwise”—naming the frameworks it does not hold; OWASP LLM Top 10 mapping names both covered and partial items; response integrity via HMAC-SHA256 signing and content hashing; zero-retention mode and metadata-only logs.
Cons: Explicitly not certified—no SOC 2 Type II, no ISO 27001, no FedRAMP, no signed BAA and therefore no HIPAA offering for the hosted service, so regulated buyers must rely on self-hosting or wait; name collision risk—the PyPI promptguard-py library and multiple Chrome extensions share the name but are unrelated products, so users may land on the wrong tool; company is a UK-registered entity (PG Tech Ltd) but funding, team size, and named customers are not disclosed; independent third-party user reviews on G2/Capterra/Trustpilot did not surface; availability target is 99.9% with no signed SLA behind it; Shadow AI fleet rollout is sold separately by seat, and Gateway plans include only two personal devices.
Who is it for?:Engineering teams deploying LLM features that need prompt-injection and PII protection before the model sees the request; security teams that require air-gapped or self-hosted deployment for data-sovereignty reasons; organisations whose staff use public AI tools and need on-device PII masking; developers already using OpenAI-compatible SDKs who want protection without rewriting provider code.