# Runtime protection for customer-facing LLM apps

URL: https://airuntimesecuritytools.com/guides/runtime-protection-for-customer-facing-llm-apps/
Updated: 2026-11-16

A public LLM app faces prompt injection, sensitive data exposure, harmful output and runaway use, all named in OWASP's 2025 list. Controls sit in the request path: input checks, output checks and data filters. Sources read on 16 Nov 2026 show cloud and edge vendors document these checks.

## What attacks hit a public LLM app?

Anyone who can type can try to subvert the model, so the first attacks are prompt injection, attempts to extract the system prompt, coaxing out sensitive data and abusing cost. OWASP's 2025 list names LLM01 prompt injection, LLM02 sensitive information, LLM07 system prompt leakage and LLM10 unbounded consumption.

Output is a risk too. OWASP names LLM05 improper output handling for responses passed on without checks.

Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/).

## Which controls sit in the request path?

Three checks sit in the path: an input check before the model, an output check before the user, and a data filter on both. Amazon Bedrock Guardrails evaluates input prompts and model completions against defined filters, and can also be called on its own through the ApplyGuardrail API.

Azure's Prompt Shields analyze text for user prompt attacks and document attacks, and in Foundry can block flagged inputs before the model processes them.

Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html); [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection).

## What can an edge or gateway layer do?

It can score and block before traffic reaches your servers. Cloudflare says AI Security for Apps scores each prompt for injection so customers can block with WAF rules, and AI Gateway Guardrails can flag or block prompts before the model.

Google's Model Armor screens prompts and responses and also works for applications on other clouds.

Sources: [Cloudflare AI Security for Apps](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/); [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview).

## How do you handle sensitive data?

Redact or block at the boundary. Bedrock's sensitive information filters can block or mask personal data in inputs and model responses, based on probabilistic detection.

Probabilistic means some misses and some false blocks, so log what is masked and sample it.

Sources: [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html).

## What should you log and limit?

Log each decision with the policy that fired, and set limits on tokens and request rates so one user cannot run up cost.

Model Armor's documentation warns that false positives can degrade the user experience by blocking legitimate prompts, so tune thresholds on real traffic before enforcing.

Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview).

## What does a minimum setup look like?

A minimum setup has an input check for prompt attacks, an output check for harmful content and sensitive data, rate and token limits, and a log of decisions. Most cloud providers cover the first two with a built-in filter.

Then add monitoring for the system prompt and tools. OWASP's LLM07 covers system prompt leakage, and its LLM06 covers tools given too much power.

Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/); [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview).

## What should you not rely on?

Do not rely on the system prompt as a security boundary, on a single filter, or on a vendor claim with no public method. Google describes false positive trade-offs for its filter settings, so each filter needs tuning on real traffic.

Layer an input check, an output check and a limit on what tools the app can call, and review blocked and allowed samples every week.

Sources: [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview).

## Which events should page a person?

Page a person for a burst of blocked prompt attacks from one source, a spike in sensitive data blocks, a control failure and any change to the policy outside a release.

Send the rest to a daily review. Keep the thresholds in the runbook, and test the alert path before launch.

## Sources

- [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) (genai.owasp.org), read 2026-11-16
- [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html) (docs.aws.amazon.com), read 2026-11-16
- [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html) (docs.aws.amazon.com), read 2026-11-16
- [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection) (learn.microsoft.com), read 2026-11-16
- [Cloudflare AI Security for Apps](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/) (developers.cloudflare.com), read 2026-11-16
- [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/) (developers.cloudflare.com), read 2026-11-16
- [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview) (cloud.google.com), read 2026-11-16

## Related pages

- [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/)
- [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/)
- [Input Filtering (glossary)](/glossary/input-filtering/)
- [All guides](/guides/)
- [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/)
- [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/)
- [The ranking](/ranking/)
- [Glossary](/glossary/)
