ARCHITECT THE
AUTONOMOUS WEB

"Lakera Guard is a detection layer, not a guarantee. It scores every request in and every response out, blocks what crosses a threshold, and is model-agnostic — it works whether the model runs on a hosted API or on your own hardware. Its filters are trained partly on data from Gandalf, Lakera's public prompt-injection game, which has collected millions of real attack attempts. Check Point acquired Lakera in 2025 (reported around $300 million); the product now also ships as Check Point AI Guardrails, while the standalone Guard API and its free tier still exist. The detection-rate, latency and language-count figures Lakera publishes (98%+, sub-50ms, 100+ languages) are the vendor's own — treat them as a claim, not a measured result. The real trade-off is architectural: you add a network hop and an external dependency to every LLM call, in exchange for a maintained, continuously updated filter you do not have to build or keep current yourself."

"Garak — "generative AI red-teaming and assessment kit" — is the closest thing the LLM world has to nmap or Metasploit, an analogy the project uses itself. You point it at a target (an OpenAI or Anthropic API, a Hugging Face model, a local Ollama or llama.cpp endpoint), it runs a battery of probes, and it produces a JSONL report marking which attacks scored a hit. It covers known jailbreak families (DAN and relatives), encoding and Unicode tricks that slip past naive filters, forced hallucination, toxicity and training-data leakage. It was created by Leon Derczynski and is now maintained by NVIDIA and the community under Apache 2.0. Two honest caveats: it finds problems, it does not fix them, and scanning a hosted model runs real inference — you pay for the tokens and can trip rate limits. Reading the output well takes some security literacy."