PodcastRe-Platforming Your Tech StackClaimThe team used parts of Vertex AI and relied heavily on Python APIs to move models while minimizing changes that would trigger additional governance work.23:28
Guardrails, governance and security in 2025
41 sessions
PodcastRe-Platforming Your Tech Stack
PodcastEvolving Workflow OrchestrationClaimDrawing workflows can reveal which systems and services an organization uses, what data moves between them, and where governance or compliance risks exist.1:06:41
PodcastAI SQL Data AnalystClaimEach analyst has development and production versions of the agent, and only trusted use cases are released to production users.38:27
AI in Production 2025 | KeynotePushed 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
PodcastBuilding Trust Through Technology: Responsible AI in PracticeClaimResponsible AI principles can include fairness, accountability, transparency, explainability, privacy, safety, reliability, and robustness.1:44
PodcastFraud Detection in the AI EraClaimRafael Sandroni says guardrails should control outputs such as profanity, competitor references, hallucinations, and off-topic responses.2:56
PodcastWe're All Finetuning IncorrectlyClaimLLMs are useful as flexible prototype components and are particularly good at fluency, but they should not be trusted as unconstrained decision-makers.37:44
Which Economic Tasks are Performed with AI? Evidence from Millions of Claude ConversationsPushed backAdam Becker questions whether removing small clusters for privacy protection could distort the results.19:26
Iceberg, MCP, and MLOps: Bridging the Gaps for EnterprisePushed backHamza Tahir argues that both rigid stack enforcement and unrestricted stack sprawl create problems, favoring a flexible governance envelope instead.24:04
AI Agents for Financial InstitutionsClaimFinancial AI agents need customizable guardrails, compliance workflows, source citations, reasoning transparency, and integrations with core banking and digital payment systems.2:29
MCP is not going to change everything (yet)ClaimTool authorization adds an agent as an intermediary between the user and a service, so the agent needs its own scopes, claims, and permissions.6:44
A New Way of Building with AIPushed backAI integrations should adapt interfaces to model behavior rather than simply impose narrow developer-designed scopes and guardrails.10:49
MLflow 3.0: The Future of AI AgentsClaimEric Peter identifies access to LLMs, end-to-end observability, integration with authoring frameworks, open standards, and enterprise security and governance as five key needs for agent operations.6:16
AI Changed Stack Overflow for the BetterPushed backPrashanth Chandrashekar reported that trust in AI output declined even as AI adoption for coding increased.17:23
How AI Will Transform The Energy SectorClaimAdam Sroka says Hypercube is also building a compliance-automation tool to test language models and monitor issues such as prompt injection and bias.21:53
APICA: The Digital Colleague at the Port of Antwerp-BrugesClaimAPICA's safety approach includes monitoring and alerting, access control, guardrails, and a security evaluation framework.7:17
Deploying Executable Agent WorkflowsClaimGal Peretz said that capable models such as the OpenAI models may work with trustcall, while open-source models may need fine-tuning to restrict their output to OpenAI-style JSON schemas.16:08
Integration of AI into Traditional SystemsPushed backHakan Tek rejects the idea that using a third-party service can provide complete certainty about data security.14:30
Building Real-Time, Reliable Voice AI: From Simulation to ProductionClaimGuardrails for real-time voice output should generally run asynchronously so they do not block the conversation, with classifiers checking the text output and cutting off unwanted content.17:55
How to Self-Host an AI AgentPushed backSelf-hosting does not eliminate data-security risk, but it can reduce the level of risk when the model runs in the same environment as the data and additional access controls are used.19:47
Zero Trust for Multi-Agent SystemsClaimApplying zero trust at the agent level means using scoped permissions, mutually authenticated communication, signed activity logs and dynamic quarantine for malicious or malfunctioning agents.5:51
PodcastIs Open Source Software Actually Secure?ClaimHudson Buzby says AI services that become core organizational services must receive the same security, governance, legal, infrastructure, and development scrutiny as other production software.10:36
Building Agents for HealthcareClaimCorti's orchestrator uses governance, auditing, and specialized experts to answer healthcare queries and request more information when needed.10:42
The Cost of AI: FinOps Strategies for Intelligent AgentsClaimAdvait Patel identifies infinite retries, repeatedly reloading full context, indefinite log retention, uncontrolled autoscaling, excessive permissions, and poor visibility as common causes of agent cost increases.3:31
What It Takes to Run Multi-Agent SystemsClaimCost and privacy and security concerns are the two main roadblocks preventing enterprises from scaling AI.3:37
How to Build Execution Layers That Don't Burn OutClaimAutonomy means taking responsibility for solving a task while still respecting guardrails and refusing inappropriate or confidential information.2:20
Why You Should Care About Observability in LLM WorkflowsPushed backColin McNamera said he was not a major fan of some of Langfuse's security architecture, particularly around prompt handling.11:58
Evals Aren't Useful? Really?ClaimMulti-turn simulations can test whether guardrails resist persistent users trying to obtain restricted information or discounts.5:17
How to build agents that take ACTIONClaimAlex Salazar identifies security, cost, latency, and accuracy as the main reasons agents fail to reach production.4:08
PodcastThe GPU Uptime BattleClaimScaling an AI application requires security, privacy, lineage, governance, and authorization for both source data and derived data.1:30:55
AI Agents Are About to Get Real, Here's Why It MattersClaimArcade helps an agent take actions on a user's behalf securely, even when the user is not near a computer.1:03
PodcastGovernance for AI Agent DeploymentClaimSpencer Reagan says governance must balance agent autonomy with constraints because agents are non-deterministic and cannot be controlled by a fully discrete rule set.27:28
PodcastThe Future of AI Agents are SandboxesClaimA sandbox gives an agent its own isolated computer where it can operate with reduced security risk and a controlled blast radius.0:47