Multimodal Risk Policy

This new policy addresses risks specific to multimodal AI systems that process or generate combinations of text, image, audio, and video, closing the gap left by the Content and Synthetic Media Policies.

All multimodal-capable models and features across AnswerAI, WSAI CODE, and WEBSPACEAI Studio.

• Cross-Modal Injection — malicious instructions embedded in a non-text modality (image, audio, video) intended to manipulate model behavior when processed.
• Audiovisual Hallucination — generated audio or video content that fabricates realistic but false depictions of real events or people.

Multimodal models must pass cross-modal jailbreak resistance testing (per Doc. 17 and Doc. 18) before release. Image and video inputs are scanned for embedded text/steganographic instructions before being passed to the model context. Audio inputs are transcribed and scanned through the same instruction-hierarchy enforcement as text (per Doc. 19). Audiovisual outputs depicting real people or events carry provenance metadata and, where realistic depiction of unverified events is detected, an in-product disclosure banner.

Research Engineering owns multimodal safety tooling; AI Safety Research Lead reviews new multimodal capabilities before release; Studio product team implements provenance and disclosure banners.

Cross-modal testing is mandatory for all Level 2+ multimodal releases; results documented in the Safety Report.

Multimodal features failing cross-modal testing are blocked from release.

safety@webspaceai.in

Requires AI Safety Research Lead approval, time-limited pending remediation.

Quarterly given the pace of multimodal threat evolution.

safety@webspaceai.in