Amazon Bedrock Guardrails launches new safety features to fight harmful generative AI content.
The tool promises up to 88% better blocking of unsafe multimodal content. It also filters over 75% of hallucinated responses in retrieval and summarization tasks. This is Amazon’s first safeguard using Automated Reasoning to catch factual errors.
The launch targets industries like healthcare insurance, where the AI assistant can manage claims, coverage checks, and provider info while keeping operations secure. Guardrails run checks at model, prompt, and application levels. 
Amazon Bedrock Guardrails works with all large language models on Amazon Bedrock and even external foundation models through the ApplyGuardrail API. It can block hate speech, insults, violence, sexual content, misconduct, and prompt injection attacks.
Setting up a guardrail involves configuring content filters, denied topics, word filters, and sensitive information blocks—like emails and phone numbers—to comply with regulations such as HIPAA.
Automated Reasoning checks are in preview in the US West (Oregon) region. It translates policy documents into logical rules that the AI must follow, flagging invalid or risky answers.
Developers can test guardrails via the console or the ApplyGuardrail API, which works independently of model invocation. This flexibility supports deploying safeguards on models hosted outside Amazon Bedrock, including on-prem or third-party models.
Amazon highlights that this framework builds trust by enforcing AI safety and compliance at scale.
AWS Solutions Architect Divya Muralidharan and Principal Technologist Rachna Chadha detail the features in a GitHub repo and a walkthrough post.
Amazon Bedrock Guardrails now offers a powerful toolset for companies handling sensitive data who want tighter control over generative AI content.
