Aporia Introduces Real-Time Guardrails for Multimodal AI Applications

Aporia Technologies Ltd., a machine learning observability startup, today announced a new service, Guardrails for Multimodal AI Applications, expanding its existing artificial intelligence guardrails solution to also include vision and voice guardrails.

The first of its kind, Guardrails for Multimodal AI Applications is designed to mitigate issues in video- and voice-based AI applications, including hallucinations, false responses, compliance violations, jailbreak attempts, etc. Multimodal AI is an AI system that can simultaneously process and interpret multiple types of data inputs, including text, images, audio, and video.

The release of this new service follows OpenAI's unveiling of its flagship multimodal GPT-4o model on May 13. Aporia claims that while GPT-4o offers unprecedented productivity with the richest, most human-like AI experience to date, it also raises major questions about accountability.

The issue comes down to what goes into an AI model and what gets returned. Aporia said misleading information to users could have serious repercussions, like an AI advising drug or alcohol abuse on someone looking for ways to cope with depression, or a banking customer asking to see their financial history only to get someone else's data in return.

Guardrails for Multimodal AI enable engineers to add an extra layer of security and control between apps and users. Guardrails work by using a set of predefined, customizable behavior rules that operate with sub-second delays, going beyond the capabilities of typical prompt engineering.

“Multimodal AI is a game-changer for the world we live in, but guardrails are needed to ensure its safety, success, and ultimate adoption,” said Liran Hasson, CEO and co-founder of Aporia. “Industries around the world have become reliant on AI, but as many engineers are finding out, AI itself is inherently untrustworthy.”

According to the company, the new service detects and mitigates 94% of hallucinations in real time before they reach the user, providing a powerful layer between large-scale language models and AI applications. The solution also prevents applications from being abused for malicious purposes, such as prompt injection and prompt leak, blocks explicit and offensive language in user interactions, and can identify and instantly block inappropriate language and phrasing.

“As we've seen, tragic accidents can happen suddenly,” Hasson added. “Aporia Guardrails is the first solution to proactively mitigate verbal and written responses in real time to support human-involved environments.”

Aporia is a venture capital-backed startup that raised $25 million in February and has raised $30 million to date, according to Tracxn. The company's investors include Tiger Global Management LLC, TLV Partners LP, Samsung NEXT LLC and Vertex Ventures.

Image: Aporia