New Features - Prisma AIRS - March 2025

Code snippets generated by Large Language Models (LLMs) can be protected with Malicious Code Detection feature for potential security threats. This feature is crucial for preventing supply chain attacks, enhancing application security, maintaining code integrity, and mitigating AI risks.

The system supports scanning for malicious code in multiple languages, including JavaScript, Python, VBScript, PowerShell, Batch, Shell, and Perl.

To activate this protection, you need to enable it within the API Security Profile. When configured, this feature can block the execution of potentially malicious code or be set to allow, depending on your security needs. This capability is vital for organizations that are increasingly leveraging generative AI for development, as it helps to secure against the risks of LLM poisoning, where adversaries intentionally introduce malicious data into training datasets to manipulate model outputs.

​​To protect AI applications from generating or responding to inappropriate content, a new capability adds toxic content detection to LLM requests and responses. This advanced detection is designed to counteract sophisticated prompt injection techniques used by malicious actors to bypass standard LLM guardrails. The feature identifies and mitigates content that contains hateful, sexual, violent, or profane themes.

This capability is vital for maintaining the ethical integrity and safety of AI applications. It helps protect brand reputation, ensures user safety, mitigates misuse, and promotes a responsible AI. By analyzing both user inputs and model outputs, the system acts as a filter to intercept requests and responses that violate predefined safety policies.

The system can either block the request entirely or rewrite the output to remove the toxic language. In addition to detecting toxic content, it also helps prevent bias and misinformation, which are common risks associated with LLMs. By implementing this security layer, you can ensure that your AI agents and applications operate securely and responsibly, safeguarding against both intentional and unintentional generation of harmful content.