All numbers are reproducible on held-out data → see the honest report
The whole agent lifecycle — not just the prompt.
Detects hidden instructions, jailbreaks, and prompt-leak attempts in content your agent reads (web, email, tool output). Rules + Prompt Guard 2 + LLM ensemble.
Catches secret keys, credentials, PII (national IDs, cards), and system-prompt disclosure before they leave — deterministic secrets + PII engines.
Tracks the session's action sequence to catch exfiltration chains, hijack execution, credential stuffing, enumeration, and bot-speed automation.
No rewrite. Change one line, or add one decorator.
# Proxy mode — zero code change, guards any OpenAI-compatible API client = OpenAI(base_url="https://api.cogitapp.com/guard/proxy/v1", api_key=KEY) # or SDK — one decorator from neosguard import Guard guard = Guard(api_key="...") @guard.protect def agent(user_text): ...
Teach it a new attack once and it becomes immune to that attack and its variants in seconds — verified, zero false positives. Live demo →
A red-team loop invents new attacks; a hard verifier adopts a defense only if it catches the attack with zero new false positives — so it strengthens monotonically, never degrades.
Attacks and leaks in English, Korean, Japanese, Chinese — not an English-only filter. Global agents face global attackers.
Run fully local (no data leaves your infra) or use the hosted API. Free to start.
No "99% accuracy" marketing. Held-out metrics, OWASP LLM Top 10 coverage, and a self-improvement log — all public.
Free to start. Scale as you grow.