AI Security
Akamai Firewall for AI
AI threats arrive in natural language, where a WAF has nothing to match on. Firewall for AI is an inline checkpoint in the request path: AAM calls it, gets a verdict back, and acts on that verdict before the request goes any further.
How it works
AAM calls Firewall for AI mid-request and waits for the answer — on the prompt going out and again on the response coming back.
Enforced in the same request. Nothing is mirrored for later analysis and no blocklist has to catch up — the verdict comes back while the caller is still waiting, and AAM acts on it before the traffic moves on. Inspection is a per-route policy, so you apply it where the risk is and leave the rest of your traffic untouched.
The decision
Every verdict maps to an action AAM takes immediately.
Clean traffic passes straight through to the model or tool.
PII, secrets, or confidential content are redacted before the payload moves on.
The request proceeds, but the verdict is logged and alerted on for review.
The request is rejected. The prompt never reaches the model.
What it detects
Prompt injection
Instructions planted in user input or tool output to hijack model behavior.
Jailbreaks
Prompts engineered to bypass safety constraints or role restrictions.
Data exfiltration
Model responses carrying PII, credentials, or confidential content.
Toxic and off-policy content
Output that breaches your content policy on the way back to the caller.
Where it runs
The same inline check attaches to any AAM traffic, across all three AI products.
Checks prompts sent by AI agents calling your APIs.
Checks prompts on the way in and model output on the way back.
Checks tool calls and tool responses before an agent can act on them.
Firewall for AI complements AAM's API-layer controls rather than replacing them. Authentication, rate limiting, and token budgets stop volumetric and access-control abuse; Firewall for AI handles what is only visible in the content of a prompt or a response.