AI-Generated Media in the Wild: Risks, Failures, and Mitigations
About This Session
As generative AI systems evolve from text to image and video, the nature of risk is shifting - becoming more complex, harder to detect, and increasingly impactful at scale.
In this session, I’ll share practical insights from building and deploying AI media safety systems at Meta, with a focus on how risks in generative systems are discovered, evaluated, and mitigated in real-world environments.
We’ll cover:
1. Risk discovery in generative media systems - identifying failure modes across images and video, including adversarial prompts and edge cases
2. Post-training alignment techniques, including reinforcement learning approaches to reduce harmful outputs
3. Automated red-teaming pipelines for systematically uncovering visual and temporal risks before deployment
4. Multimodal safety classifiers that balance risk reduction with usability (minimizing false positives and unnecessary refusals)
5. Scaling evaluation frameworks to increase testing coverage and iteration speed
6. Operationalizing safety through automation, including content risk labeling systems that significantly reduce manual effort and cost
This talk is intended for engineers, researchers, and practitioners working on deploying AI systems responsibly. The goal is to provide concrete patterns for identifying, measuring, and mitigating risk in modern AI systems - especially those operating at scale in real-world environments.
In this session, I’ll share practical insights from building and deploying AI media safety systems at Meta, with a focus on how risks in generative systems are discovered, evaluated, and mitigated in real-world environments.
We’ll cover:
1. Risk discovery in generative media systems - identifying failure modes across images and video, including adversarial prompts and edge cases
2. Post-training alignment techniques, including reinforcement learning approaches to reduce harmful outputs
3. Automated red-teaming pipelines for systematically uncovering visual and temporal risks before deployment
4. Multimodal safety classifiers that balance risk reduction with usability (minimizing false positives and unnecessary refusals)
5. Scaling evaluation frameworks to increase testing coverage and iteration speed
6. Operationalizing safety through automation, including content risk labeling systems that significantly reduce manual effort and cost
This talk is intended for engineers, researchers, and practitioners working on deploying AI systems responsibly. The goal is to provide concrete patterns for identifying, measuring, and mitigating risk in modern AI systems - especially those operating at scale in real-world environments.
Speakers
Ksheeraj Vepuri
Senior Research Engineer - Meta
Ksheeraj Vepuri is a Senior Research Engineer at Meta Superintelligence Labs, where he works on AI media safety and alignment for generative image and video systems. His work focuses on post-training alignment, automated red-teaming, and large-scale evaluation frameworks for safe deployment. Previously at Expedia Group, he led generative AI initiatives for Virtual Agents, driving LLM adoption and agentic systems in production. He is interested in building scalable AI systems that are safe, aligned, and reliable in the real world.
Sourabh Kulkarni
Research Scientist - Meta Platforms
Sourabh Kulkarni, PhD is a Research Scientist at Meta, where he leads safety research and evaluation for generative media foundation models -- image, video, and audio -- ensuring responsible deployment to billions of users worldwide. His current work spans adversarial red-teaming, multilingual safety evaluation pipelines, and policy-aligned harm mitigation for pre-launch generative systems. Previously, Sourabh was a core contributor and ML tech lead for Meta's "Why am I seeing this ad?" transparency product, the first-of-its-kind algorithmic explainability system serving ~3.5 billion users daily, and architected agentic content-generation systems behind Meta Vibes.