How Secra Works

No. The majority of attacks are caught by Layers 0 and 1 which use zero LLM calls.

Layer 0 is a pure string-matching engine built on the Aho-Corasick algorithm. It scans for 30+ injection signatures in a single pass in under 1 ms. No AI. No API call. No token cost. Ever.

Layer 1 is a rule engine using regex and heuristics. Still no LLM. Still free.

Layer 2 — the only layer that calls an LLM — only fires when the first two layers are genuinely uncertain. That is a small minority of real-world traffic.

Most attacks never reach Layer 2.

Only when the combined confidence score from Layers 0 and 1 falls between 0.25 and 0.75 — the ambiguous zone.

A direct override attack like "ignore all previous instructions" scores above 0.75 immediately at Layer 0. Groq never sees it. You are never charged for it.

Groq running llama3-8b-8192.

We chose Groq specifically for speed. Even when Layer 2 fires, response time stays between 50–200 ms — fast enough that your users won't notice.

Groq's inference speed means the security check doesn't become a bottleneck in your agent's response time.

Layers 0 and 1 always run regardless of Groq's status. The vast majority of attacks are caught at these layers.

For ambiguous prompts during a Groq outage, Secra defaults to a configurable fallback behaviour:

Enterprise plans include SLA guarantees and dedicated fallback configuration. Set your preference in your dashboard.

For context — a typical LLM call to GPT-4 or Claude takes 500 ms to 3 seconds. Secra's pre-processing adds negligible latency for the vast majority of requests.

Your users will not notice Secra is there. That is the point.

Block Mode stops the request entirely and returns a 403.

Sanitize Mode strips the injection payload from the prompt and returns a clean version your LLM can safely process. The user's legitimate intent is preserved.

Your LLM receives the safe version. Your user gets their answer. The attack is silently removed.

Use Sanitize Mode when blocking would break your user experience.

AI agents that call tools — APIs, functions, MCP servers — are vulnerable to function injection attacks. An attacker crafts input that causes your agent to call unintended tools with malicious arguments.

Secra inspects the tool name and arguments before execution. If the call looks like an injection attempt it is blocked before your tool runs.

This protects the action layer of your agent, not just the input layer.

Your sk_secra_ keys are:

Never hardcode keys in your source code. Use environment variables.

Your dashboard shows: verdict (blocked or passed), threat category, which layer caught it, latency, and token spend.

We do not store the full content of passed prompts. Blocked prompts are logged for your security audit trail and automatically purged after 30 days.

Enterprise plans include custom data retention policies.

Tokens are only charged when Layer 2 fires — the Groq LLM call.

Layer 0 and Layer 1 scans are always free regardless of your plan.

Tokens reset monthly. No seat fees. No setup costs. No credit card on the free plan.

Tested integrations:

If your stack is not listed, the REST API works with any HTTP client. Open a GitHub issue and we will add official support.