The AI Firewall operates as a transparent proxy. You point your AI SDK at Kovral's endpoint — one configuration change — and every interaction is inspected without modifying your application code.
Your application sends a prompt to, say, OpenAI's API. The request arrives at Kovral's regional inspection endpoint co-located with the AI provider.
Kovral's inspection engine runs the prompt through 10,000+ threat signatures, semantic similarity matching, and ML-based anomaly detection — simultaneously, in under 2ms.
The prompt is evaluated against your active policies: data classification rules, acceptable use policies, user-group restrictions, and model-specific rules.
If the prompt passes: it's forwarded to the AI provider with a risk score attached. If it violates policy: it's blocked (a safe decline message returned to the user) or redacted (sensitive data masked, then forwarded).
The model response is inspected for exfiltration indicators, PII leakage, jailbreak success signals, and output policy violations before being returned to your application.
Kovral's threat signatures cover the OWASP Top 10 for LLMs plus proprietary threat classes discovered through our research programme.
Direct and indirect injection attempts including system prompt override, jailbreak prefix attacks, role-playing exploits, and multi-step context manipulation chains.
Prompts designed to extract training data, system prompts, or other users' data from the model. Response scanning catches models that inadvertently leak sensitive data.
PII (names, emails, SSNs, credit cards), API keys, passwords, and source code in both prompt inputs and model outputs — detected with LLM-aware pattern matching.
Known jailbreak techniques including DAN, role-play exploits, base64 encoding attacks, and semantic evasion variants. Updated daily from our red team research.
Malicious instructions embedded in documents, web pages, or tool outputs that attempt to override an autonomous AI agent's objectives or exfiltrate data through its tool calls.
Unusually long prompts, recursive generation requests, and crafted inputs designed to exhaust token budgets or trigger excessive compute on the model.
