Opencontroller by Lyzr
Opencontroller blog
Answers to the questions enterprises ask about governing AI agents
Discovery and shadow agents
All in this topicHow do you find AI steps hidden inside Zapier, Make and Power Automate workflows?
Find AI steps in no-code workflows: Power Platform inventory, Make usedPackages and blueprints, Zapier app connections, and the HTTP steps most scans miss.
Why does an AI agent inventory show duplicate or mismatched records, and how do you fix it?
Duplicate agent records come from two sources seeing one agent, versions counted as agents and renamed resources. How to pick matching keys and reconcile.
Agent identity and access
All in this topicRuntime control
All in this topicTools and MCP
All in this topicHow do you run a private MCP registry for your enterprise?
Build a private MCP subregistry: mirror approved servers from the official registry, add approval fields, pin versions and point clients at it.
What are MCP sampling and elicitation, and are they safe?
MCP sampling lets servers request model completions; elicitation lets them ask users for input. How each works, the risks, and client-side controls.
Gateways and cost
All in this topicHow do you keep your prompts, evaluations and logs when you switch model providers?
Own the prompts, eval sets, traces and routing rules that make your agents work, and avoid depending on state the model provider holds for you.
How do you measure productivity when several AI tools contribute to the same outcome?
When assistants and agents touch the same work, vendor dashboards double count. Measure outcomes at the process level with cohorts and holdouts.
What does a CISO security review of an LLM gateway cover?
The nine areas a CISO checks before an LLM gateway goes live, from SSO-only admin access to prompt logging and egress IPs, with pass criteria for each.
How do you separate AI coding assistant costs from AI agent runtime costs?
Build costs (coding assistants, premium requests, CI agents) and runtime costs (tokens per agent run) have different drivers. A tagging scheme to split them.
Evidence and frameworks
All in this topicWhat do California's ADMT rules mean for AI agents that make significant decisions?
California's ADMT rules apply from 1 Jan 2027. When an AI agent counts as ADMT, which decisions trigger it, and what evidence to keep.
What does the NIST Cyber AI Profile mean for securing AI agents?
The draft NIST Cyber AI Profile (NIST IR 8596) applies CSF 2.0 to AI. Which CSF categories matter most for AI agents, and what to evidence.
How do you score AI agent vulnerabilities with OWASP AIVSS?
OWASP AIVSS starts from a CVSS v4.0 base score and adds an agentic uplift from 10 amplification factors. How to score an agent vulnerability, step by step.
What evidence does ETSI EN 304 223 expect for securing AI agents?
ETSI EN 304 223 sets 13 baseline security principles across five AI lifecycle phases. Each principle mapped to evidence you can produce for AI agents.
What evidence do the MAS AI risk management guidelines expect for AI agents?
MAS published AI risk management guidelines on 7 Oct 2026, effective 7 Oct 2027. What financial institutions should evidence for AI agents.
When does building an AI agent make you a provider under the EU AI Act?
Build an agent on a third-party model and run it under your name, and you are usually its provider. Article 25 triggers, worked examples, evidence.
Does the EU AI Act still require AI literacy for people who use AI agents?
Yes. Article 4 applies since Feb 2025 and the Digital Omnibus softened but kept it. A checklist of AI literacy evidence for staff who use or run AI agents.
What do the EU AI Act Digital Omnibus changes mean for AI agent deadlines?
Regulation (EU) 2026/1744 moved high-risk deadlines to Dec 2027 and Aug 2028. What changed for AI agents, what did not, and what evidence to build now.
What counts as a serious incident under the EU AI Act, and when must you report it?
Article 73 sets 15-, 10- and 2-day deadlines for serious AI incidents. Definitions, who reports, how the 2026 omnibus affects timing, and logging.
How do you run a GDPR DPIA for an AI agent?
When GDPR Article 35 requires a DPIA for an AI agent, the WP248 criteria agents meet, and what each DPIA section must cover for agent data flows.
What PCI DSS evidence do you need when AI agents touch card data?
Agents that see card data pull models, gateways, memory and logs into PCI scope. The v4.0.1 requirements that change and the evidence to collect.
How does the CSA AI Controls Matrix apply to AI agents?
The CSA AICM v1.1 has 247 controls in 18 domains. Which domains cover AI agent risks, who owns each control, and the evidence to collect.
What are NIST control overlays for AI systems and agents?
NIST is tailoring SP 800-53 for five AI use cases, including single- and multi-agent systems. Project status and the controls to map today.
What does the Colorado AI Act mean for AI agents?
Colorado replaced SB 24-205 with SB 26-189, effective January 1, 2027. What changed, when agents count as ADMT, and the notices and records to prepare.
Buying and comparison
All in this topicWhat is the total cost of ownership of an AI agent governance platform?
The licence is one line of six. A worksheet for the real cost of AI agent governance: infrastructure, integration, people, gateway compute and exit costs.
How do you score an AI agent build platform for governance readiness?
A weighted 100-point rubric for scoring agent build platforms on governance: identity, traces, kill switch, budgets, external agents, deployment and evidence.
Can you run an AI governance proof of concept on dummy data?
What a dummy-data proof of concept for AI agent governance can and cannot prove, and a staged plan from synthetic data to read-only production connectors.
Platform Guides
All in this topicWhich authentication setting should Copilot Studio agents use?
No authentication vs Authenticate with Microsoft vs manual: which to use, why anonymous agents are a red flag, and the admin controls that block them.
How do you trace Microsoft Foundry agents with Application Insights?
Foundry traces agents with OpenTelemetry into Application Insights. Setup steps, the spans you get, and the content-recording setting to govern.
How should AI agent pods on Kubernetes get cloud credentials?
One service account per agent, bound with EKS Pod Identity or IRSA, GKE Workload Identity Federation or AKS Entra Workload ID. No static keys.
Which GitHub audit log events show Copilot policy and agent activity?
Search action:copilot for plan events and actor:Copilot for agent actions. Key fields, the 180-day limit, SIEM streaming and what the log misses.
How do you use AgentCore Observability for AI agent audit trails?
AgentCore Observability sends OpenTelemetry spans to CloudWatch. Where the data lands, how to keep it, and what an audit trail still needs.
How do you control retention and PII in Amazon Bedrock AgentCore Memory?
AgentCore Memory retention: event expiry up to 365 days, long-term record TTLs, namespaces per user, KMS keys and how to erase a user's memories.
How does Claude Code sandboxing limit what a coding agent can touch?
Claude Code sandboxing uses OS-level isolation to limit shell command writes, reads and network access. How to enforce it org-wide and its limits.
How do admins control connectors and custom MCP servers in ChatGPT Enterprise?
ChatGPT Enterprise admins control apps (connectors), action permissions, developer mode and custom MCP servers through workspace settings and RBAC.
How do you govern Vertex AI Agent Engine Memory Bank?
Memory Bank scopes, TTLs, revisions, purge and IAM Conditions: how to control what Google agent memory stores and how to delete it on request.
How do you use Anthropic Console workspaces to cap agent spend?
Anthropic workspaces let you set monthly spend limits, per-model rate limits and workspace-scoped API keys for each agent or team. Here is how.
How do you add human approval steps to LangGraph agents?
Use LangGraph interrupt() with a durable checkpointer and thread_id to pause before risky tool calls, resume with Command, and log who approved.
How does Amazon Bedrock AgentCore Runtime isolate agent sessions?
AgentCore Runtime gives each session its own microVM with idle and max lifetime timers up to 8 hours. What that protects, and what it does not.
How do you restrict models and rate limits per OpenAI API project?
Use OpenAI API projects to allow only approved models, set per-model rate limits, budgets and service account keys for each agent or team.
How do you run Microsoft Foundry agents on a private network?
Foundry Agent Service standard setup with your own VNet: subnet delegation and sizing, private endpoints for Storage, AI Search and Cosmos DB, egress and tool limits.
How do you stop Vertex AI agents exfiltrating data with VPC Service Controls?
A VPC Service Controls perimeter blocks public egress from Vertex AI agents. How to add controlled egress with PSC interface, roll out in dry run, and what it misses.
What permissions does a Salesforce Agentforce agent run with?
Agentforce employee agents run as the signed-in user; service agents run as a dedicated agent user. Permission set design, least privilege and audit.
How do you audit Copilot Studio agent activity in Microsoft Purview?
Which Copilot Studio maker and runtime events reach the Purview unified audit log, how to search them, retention, and the logging gaps to cover.
How do you inventory every Copilot Studio agent across Power Platform environments?
Power Platform inventory lists Copilot Studio and Agent Builder agents tenant-wide. Roles, fields, API export, refresh time and the gaps to plan for.
Do Snowflake masking and row access policies apply to Cortex Agents?
Cortex Agents run with the caller's roles, so Analyst SQL honours masking. Cortex Search runs with owner's rights. Where policies apply and how to test.
How does on-behalf-of-user authentication work for Databricks agents?
Databricks agents can act with the end user's Unity Catalog permissions instead of a shared service principal. How scopes, tokens and fallback work.
Can MLflow tracing serve as an audit trail for Databricks agents?
MLflow 3 traces record every step a Databricks agent takes. Here is what they capture, where they live, and what an auditor will still ask for.
Why do Amazon Bedrock calls fail through an LLM gateway, and how do you fix them?
Most Bedrock-through-gateway failures come from auth (SigV4 vs API keys), request format (Converse, Invoke, OpenAI-compatible) or inference profile IDs.
How do you track GitHub Copilot usage when your company has several GitHub enterprise accounts?
Copilot metrics are reported per enterprise and organisation. How to find every enterprise, grant access in each and reconcile seats and usage.
Put the framework into practice