Guardrails, governance and security
What data can a model use, and what should it be allowed to do? Privacy and governance discussions meet newer problems with prompt injection and agent permissions.
What Does Best in Class AI/ML Governance Look Like in Fin Services?
Pushed backCharles Radclyffe disputes the view that regulation should be the only reason organizations introduce AI governance controls.21:20
Build vs Buy an ML Platform
Pushed backDiego Oppenheimer argues that replacing an existing system is not automatically valuable if it is working and meeting required security, compliance, upgrade, and delivery needs.21:00
Operationalize Open Source Models with SAS Open Model Manager
Pushed backIvan Nardini agrees with Demetrios Brinkmann that security is one of the hardest parts of building an open-source end-to-end machine-learning product.17:03
What are regulations saying about data privacy?
Pushed backDemetrios Brinkmann suggests that privacy regulations may be taken less seriously by smaller and disruptive companies, and Cat Coode agrees that this is often the case.8:08
The revolution of Federated Learning
Pushed backRamen Dutta argued that raw data access is not necessary for useful analysis because statistical analysis from privacy-preserving tools can provide the needed information.18:10
Enterprise Security and Governance MLOps
Pushed backDiego Oppenheimer argues that security investment should be a conscious risk-reward decision rather than an identical day-one requirement for every organization.35:09
Machine Learning in Cybersecurity
Pushed backMonika argued that machine learning cannot completely replace cybersecurity experts, while Demetrios Brinkmann questioned whether new attacks would be missed by the models.18:21
Building Machine Learning Models into Docker Images
Pushed backLuke Marsden argued that building containers inside containers with a mounted Docker socket is a security problem.49:21
Federated Learning: Machine Learning on the Edge
Pushed backThe speaker challenged the assumption that conventional cloud-based machine learning pipelines are sufficient, arguing that federated learning is needed for personalization, privacy, security, and cost reasons.16:44
FLOps with Scaleout's Open-core Platform
Pushed backFederated learning is not automatically the answer to privacy challenges when data cannot be standardized or even a small sample cannot be shared centrally.37:51
Trustworthy Machine Learning
Pushed backKrishnaram Kenthapadi questioned whether machine learning systems are different from complex machines and medicines that people trust without understanding how they work.9:07
Building Threat Detection Systems: An MLE's Perspective
Pushed backJeremy Jordan clarified that rules and machine learning usually work together in cybersecurity rather than rules simply being replaced by machine learning.8:41
Harnessing MLOps in Finance
Pushed backMichelle Marie Conway rejects the idea that banking's security controls are merely inconvenient, arguing that strict controls are necessary to protect customers and their finances.29:40
LLM Security
Pushed backWhether current LLM engineering projects are ready for production use was challenged by the speakers' shared view that most are still demos and need stronger security guardrails.52:04
Product Strategy for LLM Features When LLMs Aren't Your Product
Pushed backFor non-user-facing cybersecurity use cases, LLM risks can be reduced with additional controls and testing, rather than treating the LLM as an unguarded production component.12:27
Designing for Forward Compatibility in Gen AI
Pushed backRohit Agarwal says current vector databases do not sufficiently enforce permissions, creating a risk that prompt engineering could retrieve unauthorized data.42:32
A Decade of AI Safety and Trust
Pushed backPetar argued that robustness verification cannot by itself establish that a machine-learning model is trustworthy because data and labeling problems must be addressed first.26:16
AI in Healthcare
Pushed backDemetrios Brinkmann asked whether using multiple guardrail systems would be redundant overkill, while Eric Landry described using Ragas for evaluation and NeMo Guardrails for runtime protections and intent-based responses.41:36
A Blueprint for Scalable & Reliable Enterprise AI/ML Systems
Pushed backNitin Aggarwal said there is no single established framework that serves as a universal or golden standard for evaluating and governing AI systems.32:52
AI-Driven Code: Navigating Due Diligence & Transparency in MLOps
Pushed backMatt disputes the idea that every codebase should have zero security warnings, arguing that the appropriate amount of technical and security debt depends on the company’s size and stage.25:55
AI in Production 2025 | Keynote
Pushed backThe speaker rejected the assumption that one universal guardrail model is sufficient and argued that guardrails should usually be specific to the concern or application.26:38
Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations
Pushed backAdam Becker questions whether removing small clusters for privacy protection could distort the results.19:26
Iceberg, MCP, and MLOps: Bridging the Gaps for Enterprise
Pushed backHamza Tahir argues that both rigid stack enforcement and unrestricted stack sprawl create problems, favoring a flexible governance envelope instead.24:04
A New Way of Building with AI
Pushed backAI integrations should adapt interfaces to model behavior rather than simply impose narrow developer-designed scopes and guardrails.10:49
Stop Building AI Like Traditional Software
Pushed backAishwarya Naresh Reganti said prompting should be a last resort for controlling agent access, preferring deterministic guardrails and rule-based access controls.16:30
Tool Calling
Pushed backLuciana Ledesma argues against uniform governance and favors governance that scales to risk through graded intervention.28:02
Time to become a hacker
Pushed backMatt Sharp disputes the assumption that the main opportunity in AI is only generative AI investment, saying cybersecurity was the leading organizational investment in 2025.12:14
MLflow Leading Open Source
Pushed backRedacting all personal information is not always the right governance goal because it can remove valuable information needed for debugging and analysis.52:40