Amazon Bedrock Guardrails now extends protection to the code domain, safeguarding user prompts, comments, variables, function names, and string literals from undesirable content.

The system strives to balance strong security controls with maintaining developer productivity in AI-assisted software development.

Key configuration steps include enabling content filters, selecting harm-related categories, setting thresholds and actions, and testing guardrails with example prompts to detect misconduct such as keylogger generation.

Bedrock Guardrails offers six core safeguards—content filters, denied topics, word filters, sensitive information filters, contextual grounding checks, and Automated Reasoning checks—configurable via the ApplyGuardrail API and compatible with Bedrock foundation models or in-app interventions.

Denied Topics filters let organizations define topic names, definitions, and optional example phrases to restrict specific coding areas, with testing available through the ApplyGuardrail API.

Guidance covers configuring code-domain safeguards, testing guardrails, and deploying them across regions with appropriate profiles and cross-region inference (CRIS) requirements.

The article outlines common prompt attack types—jailbreaks, prompt injections, and prompt leakage—and shows how Bedrock Guardrails detects and blocks such attacks in code-domain scenarios.

The PII filter has been enhanced to protect sensitive information across coding-related text, programming language code, and hybrid content, offering 31 predefined PII types plus custom regex patterns.

Author bios highlight AWS AI scientists and product leads behind Bedrock Guardrails, emphasizing the industry focus on responsible AI and safety in code-generation contexts.

Code-domain protection supports the Standard Tier and multiple programming languages (Python, JavaScript, Typescript, Java, C#, C++, PHP, Shell, HTML, SQL, C, Go) to address security risks in coding agents, including prompt injections, data exfiltration, and malicious code generation.