What is AI Guardrails

Risk detection and content labeling

• Content compliance detection — Reviews the text that generative AI receives and produces across multiple compliance dimensions, covering risk categories such as politically sensitive content, pornography and vulgarity, bias and discrimination, and harmful values. This feature helps AI-generated content comply with laws, regulations, and platform rules. Scenarios: chatbots, AI in education, intelligent customer service, and AI-generated content (AIGC) creation platforms.

• Sensitive content detection — Detects the private data and sensitive information that may leak during AI interactions, and identifies sensitive content such as personal privacy and enterprise privacy. This feature prevents training data leaks and conversational data exposure. Scenarios: AI in healthcare, AI-powered financial services, and enterprise knowledge base Q&A.

• Prompt injection detection — Defends against the injection attacks that target generative AI. This feature identifies adversarial behavior such as jailbreak instructions, role assumption inducements, and system instruction tampering, and protects AI systems against adversarial inputs. Scenarios: securing the instruction interactions of AI Agents, defending open-domain dialogue systems against adversarial attacks, and managing permissions for third-party plug-in calls.

• Malicious file detection — Parses common document formats that users upload, such as PDF, PPT, and DOC, to identify hidden executable scripts, macro viruses, nested attack code, and other malicious content. This feature prevents attackers from using file injection to take unauthorized control of AI systems or exfiltrate data. Scenarios: AI use cases that support document uploads, such as intelligent resume parsing, contract Q&A, and enterprise knowledge base construction.

• Malicious URL detection — Analyzes the links that AI receives or generates during interactions in real time, and identifies high-risk URLs such as phishing websites, malicious redirects, and links that hide attack payloads. This feature prevents foundation models from being induced to access illegal resources or becoming a vector for cyberattacks. Scenarios: AI-powered search, web page summarization, retrieval-augmented generation (RAG)-based knowledge retrieval, and automated external operations.

• Digital watermarking — In accordance with the Measures for Labeling Artificial Intelligence-Generated Synthetic Content, automatically embeds visible or invisible watermarks into AI-generated images. This feature makes AIGC content sources traceable and responsibilities attributable, and helps prevent the spread of false information and copyright disputes. Scenarios: compliance-sensitive use cases such as AIGC creation platforms, news media, government communications, and educational content generation.

Custom configuration

You can customize risk detection items, risk thresholds, and filter words in Mitigation Settings. Log on to the AI Guardrails console to enable or disable specific risk detection items at any time and build the risk detection template that best fits your business.

• Custom detection items — Configure the granular tags used in content compliance detection.

• Custom risk thresholds — Configure the threshold at which each granular tag triggers a detection hit. Thresholds are based on the model confidence score from 0 to 100 and increase in increments of 1.

• Custom filter words — Configure the sensitive words to detect and block, such as competitor names. You can add, delete, and modify words in the word library.

Which regions and service nodes does AI Guardrails support?

The console currently supports the China East 2 (Shanghai), China (Hong Kong), Singapore, Germany (Frankfurt), and US (Virginia) regions.

For the access regions supported by the backend service, see Multimodal API integration guide.

What is the detection latency of the synchronous and asynchronous detection APIs of AI Guardrails?

• Synchronous detection APIs generally respond within seconds. Run an actual API call test to get accurate latency data.

• Asynchronous detection APIs generally have a latency of about 30 seconds.

Why are detection results inaccurate, and how do I improve detection?

• For cases in which prohibited content embedded in the body text is missed, the detection policy has been adjusted. Provide the specific requestID and run the test again to verify the detection result.

• Image moderation (psDetector) may produce false positives or false negatives. Model recognition capabilities and algorithm limitations mean that 100% accuracy cannot be guaranteed. Report the specific problem images so that the detection algorithm can be optimized.

• If plain text that contains sensitive words is not identified but the dubbed audio of the same content is identified, the issue is caused by differences in multi-modal detection. Submit a ticket and provide a sample for investigation and escalation.