Every scanner on the market finds vulnerabilities after the code exists. This gate stops them before the AI writes a single line, then hands the shipped code to Moolé's scanner. Two points in the lifecycle. One platform.
Ask any AI agent for an endpoint and it reaches for the simplest pattern. Simpler is what gets exploited. Here it is fetching an invoice by ID, with no ownership check. Any logged-in user can read anyone's invoice. That is OWASP A01, the number-one risk on the internet.
The gate sits inside the developer's AI tool. The moment the model is about to write access-sensitive code, it intercepts, checks the pattern against OWASP, and issues a verdict. A vulnerability never reaches the file.
A common question: doesn't AWS Bedrock already guard this? No. Bedrock guards the model's words. The gate guards the code the model writes. Moolé scans the code that shipped. Watch a vulnerable change try to pass through, it is caught at authorship, while safe code flows clean to production.
Filters the model's input and output: harmful content, PII, prompt-injection, hallucination grounding.
Ten OWASP gates intercept vulnerable patterns before the AI commits them. The unclaimed layer nothing else covers.
Signal-first, exploitability-ranked scanning of the built artifact. Catches what slipped through and legacy code.
Bedrock guards the model's words.
Moolé guards the code the model writes.
The direct answer to the question: here is everything AWS Bedrock Guardrails actually enforces. All six policies inspect what goes into or comes out of the model. None of them inspects the application code the model writes.
Detects and blocks hate, insults, sexual, violence, misconduct, and prompt-attack content, across text and images.
guards: input + outputBlocks conversation subjects you define as off-limits, e.g. "never give investment advice."
guards: input + outputBlocks or masks 50+ PII entity types, SSN, date of birth, address, in prompts and responses.
guards: input + outputCustom blocklists of specific words or phrases you never want the model to emit.
guards: input + outputFlags hallucinations, responses not grounded in the retrieved source or irrelevant to the query. Built for RAG.
guards: outputFormal-logic validation of factual accuracy against defined rules. The only guardrail that uses formal logic.
guards: output# This line passes EVERY Bedrock Guardrail: # ✓ no harmful content ✓ no PII leaked # ✓ no prompt injection ✓ fully grounded query = f"SELECT * FROM accounts WHERE id = {user_id}" # ← and it is a SQL injection. # Bedrock never reads the code. The gate does.
The pre-generation gate is not a bolt-on. It shares Moolé's policy definitions and reports into the same governance dashboard, so a security rule written once is enforced at authorship, in CI, and on the shipped artifact.