Rakshak|
A firewall for LLMs.
Sits between your users and your model.
Blocks what goes in. Sanitizes what comes out.
LLMs in production are vulnerable. A single prompt can extract your system instructions, leak API keys, expose internal URLs, or override your model's behavior entirely. Most teams discover this during an incident — a user screenshots your system prompt, a researcher publishes your internals, or an attacker exfiltrates data through the chat window. By then it's too late.
Intercept. Classify. Block.
Intercepts every prompt before it reaches your model. Runs through three stages: pattern detection for known attack signatures, semantic similarity against a threat dataset, and an LLM classifier for novel attacks. If flagged — blocked. Model never sees it.
Regex + signature matching against known injection patterns, jailbreak templates, and encoded attack payloads.
Embedding-based comparison against a curated threat dataset. Catches paraphrased and structurally similar attacks.
Final pass through a fine-tuned classifier for novel and zero-day attack patterns. High confidence required to pass.
prompt
detect · classify · block
protected
scan · redact · sanitize
clean response
Scan. Redact. Deliver clean.
Scans every response before it reaches your user. Detects system prompt leakage, PII exposure, and policy violations. Redacts sensitive data in-place or blocks entirely depending on severity. User gets a clean response or a generic error — never the leak.
Identifies system prompt fragments, internal instructions, and configuration data in model responses.
Detects and redacts email addresses, phone numbers, names, credentials, and sensitive identifiers in-place.
Custom rules per deployment. Block, redact, or flag based on your content policy and risk tolerance.
Let the output speak.
Real runs. Real attacks. Blocked at confidence 1.0.
› Input received.
› Stage 1: pattern... MATCH
› Stage 2: semantic... MATCH
› Stage 3: classifier...
BLOCKED
confidence: 1.0
threat: prompt_injection
› Response received.
› PII scan... DETECTED
› email → [REDACTED]
› phone → [REDACTED]
SANITIZED
fields_redacted: 2
action: in-place redact
› Scanning response...
› Leakage check... DETECTED
› type: system_prompt_fragment
› severity: critical
BLOCKED
user_sees: generic error
leak_contained: true
Threat coverage.
Rakshak is in open beta.
General use case supported. Domain-specific versions — finance, healthcare, legal — in progress.
One API call between your users and your model.
Request Early Access →