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 run a shadow AI amnesty for unapproved agents?
A step-by-step shadow AI amnesty: set a window, collect agents with a short form, triage into tiers, and enforce after the deadline without punishing teams.
How do you find dormant and orphaned AI agents?
Signals, thresholds and a retire-or-reassign flow for AI agents nobody uses or owns: idle invocations, unused keys, disabled owners and stale versions.
How do you track AI agents in a Backstage software catalog?
Map AI agents, models, tools and MCP servers to Backstage entity kinds, owners and annotations, and know what a developer portal cannot see.
How do you find AI desktop apps on employee laptops with MDM inventory?
Use Intune, Jamf or other MDM software inventory to find AI assistants, coding agents and local model runners on laptops, then decide what to allow.
Agent identity and access
All in this topicShould AI agents ever get break-glass access?
AI agents should not hold or trigger break-glass access themselves. Keep emergency access human-held, time-boxed and reviewed. A checklist for agent incidents.
How do you track delegation chains when one AI agent calls another?
When agent A calls agent B for a user, keep the user, each actor and narrowing scopes in the token and the trace. A design for multi-hop delegation.
Can SPIFFE and SPIRE give AI agents workload identity?
SPIFFE and SPIRE can give AI agent workloads a verifiable, short-lived identity. Here is what they cover, how attestation works, and what they leave open.
What identity should an AI agent use when no user is present?
Scheduled and event-driven AI agents need their own workload identity, a named human owner and narrow scopes. A decision table for autonomous vs delegated modes.
Runtime control
All in this topicHow do you apply guardrails to streaming LLM responses?
Streaming lets content reach users before a guardrail check ends. Buffered vs asynchronous modes, real vendor behaviour, and when to turn streaming off.
How do you run a new AI agent in shadow mode before it acts?
Shadow mode lets a new AI agent see real inputs while its write actions are stubbed. How to design it, compare it with human outcomes, and set exit criteria.
Should AI agents follow change freezes?
Prompts, models, tools and policies are all changes to an AI agent. A freeze scope table, plus how to handle provider model retirements and emergencies.
How do you scan AI agent outputs for sensitive data before they leave?
Output-side DLP for AI agents: which detectors to use, where to place the scan, when to redact versus block, and how to handle false positives.
Tools and MCP
All in this topicGateways and cost
All in this topicHow do you attribute Microsoft Copilot agent costs to teams and cost centres?
Use Copilot Credit spending policies, the Consumption view and the Cost Management API to attribute Copilot agent costs to teams and cost centres.
When is self-hosting an LLM cheaper than an API for AI agents?
Self-hosting an LLM beats a per-token API only at high, steady GPU utilisation. Cost components, a break-even formula, and the governance needed either way.
Should AI budgets be set per user, per agent or per team?
Set AI budgets per agent first, roll them up to teams, and add per-user caps only where people drive spend. A comparison of budget scopes and how to layer them.
Why do multi-agent systems use so many more tokens?
Anthropic reported its multi-agent research system used about 15x the tokens of chat. Where the tokens go, when it is worth it, and how to control cost.
What is a denial-of-wallet attack on AI agents?
A denial-of-wallet attack runs up your model and tool bill through an AI agent. Attack patterns, the OWASP reference, and the gateway controls that stop them.
Evidence and frameworks
All in this topicHow does the new EU Product Liability Directive affect AI agents?
The new EU Product Liability Directive treats AI software as a product from 9 Dec 2026. Who is liable for AI agents, and the evidence to keep.
How does NIST AI 600-1 apply to AI agents?
NIST AI 600-1 lists 12 generative AI risks. A table mapping each to AI agent controls and evidence, plus the risks that matter most for autonomous agents.
Do China's AI content labelling rules apply to AI agents?
China's AI labelling rules (from 1 Sept 2025) require explicit and metadata labels on generated content. Which AI agents are in scope and how to implement it.
What do India's AI Governance Guidelines mean for AI agents?
India's Nov 2025 AI Governance Guidelines are non-binding. The seven sutras, six pillars, and how to map them to AI agent controls and evidence.
How do you use ISO/IEC 42005 for AI agent impact assessments?
Use ISO/IEC 42005:2025 as the method behind ISO 42001 clause 6.1.4 and A.5, adding agent inputs like autonomy, tools, identity and kill plans.
Does the EU GPAI Code of Practice apply to companies building AI agents?
Agent builders are rarely GPAI model providers. The one-third compute test, what the Code covers, and the Annex XII documents to request from model providers.
What does FedRAMP mean for AI agents used by US federal agencies?
FedRAMP authorises cloud services, not agents. What the AI prioritisation and 20x mean for agencies, plus the agent evidence to map to NIST 800-53.
What does OMB M-25-21 require for AI agents in US federal agencies?
OMB M-25-21 sets use case inventories, minimum risk practices for high-impact AI and a waiver process. How they apply to AI agents and the evidence to keep.
What does the HITRUST AI security assessment cover for AI agents?
Who can get HITRUST AI Security Certification, what its AI-specific requirements cover, how it attaches to e1, i1 or r2, and the agent evidence for each area.
What does APRA CPS 230 mean for AI agents in Australian financial services?
CPS 230 has applied since 1 July 2025. How critical operations, tolerance levels, material service providers and 72-hour notification apply to AI agents.
What does Japan's AI Promotion Act mean for companies using AI agents?
Japan's 2025 AI Promotion Act sets no fines or agent-specific duties. What it does require, how the AI Basic Plan and guidelines apply, and what evidence to keep.
What happens in an ISO/IEC 42001 certification audit?
Stage 1 and stage 2 audits, what ISO/IEC 42006 requires of certification bodies, and the AI agent evidence auditors sample.
What does the EU Cyber Resilience Act mean for AI agent software?
CRA scope for agent products, the reporting duties live since 11 September 2026, and what applies from 11 December 2027.
What does NIS2 mean for AI agents in essential and important entities?
NIS2 Article 21 measures cover AI agents in essential and important entities; Article 23 sets 24-hour, 72-hour and one-month incident reports.
Platform Guides
All in this topicHow does real-time protection for Copilot Studio agents work?
Copilot Studio can call an external threat detection service before every tool call. How the webhook works, the 1-second timeout, fail-open default and setup.
How do you control which AI coding extensions developers can install in VS Code?
Use the VS Code extensions.allowed setting and AllowedExtensions policy to allow-list approved AI coding extensions, with syntax examples and rollout steps.
How do you govern Claude Code running in GitHub Actions?
Controls for claude-code-action in CI: who can trigger runs, GitHub App permissions, OIDC instead of stored keys, tool allow-lists, turn caps and cost.
What does ChatGPT Enterprise data residency cover for agents, connectors and MCP?
What ChatGPT Enterprise data residency keeps in-region, what leaves via apps, MCP and web search, and a checklist for agent and connector use.
How do you get zero data retention for AI agents on Google Vertex AI?
The settings to switch off and features to avoid for zero data retention on Vertex AI (now Gemini Enterprise Agent Platform), as a checklist for agents.
What do Salesforce connected app restrictions mean for AI agents that call Salesforce?
How Salesforce's uninstalled connected app block, device flow ban and move to External Client Apps affect AI agents that call Salesforce APIs.
How do you stop Microsoft 365 Copilot from processing sensitive files with Purview DLP?
What the Purview DLP Copilot policy location can block (labelled files, sensitive prompts, web search, external email), its limits and a rollout plan.
What Anthropic Admin API access does an AI governance tool need, and what can it not see?
Which Anthropic Admin API credential a governance scan needs, what it returns about users, workspaces and keys, and why groups may be missing.
What does the AI Red Teaming Agent in Microsoft Foundry test?
Risk categories, PyRIT attack strategies and attack success rate in Foundry's AI Red Teaming Agent, plus its gaps for multi-turn agent abuse.
How do you require approval before AI apps are installed in Slack or Teams?
Turn on Slack pre-approved apps and Teams app permission policies or app centric management, then review AI apps against seven criteria.
Should you govern AGENTS.md, CLAUDE.md and Copilot instructions files?
Which instruction files each coding agent reads, the risks they carry, and a CODEOWNERS-based review checklist.
How do you govern Gemini in Google Workspace?
Combine per-service Gemini toggles by OU or group, the Gemini app toggle, Gemini audit log events and Drive sharing controls.
How do you use Google ADK callbacks as guardrails?
Use before_model and before_tool callbacks to block, after callbacks to redact, and ADK Plugins or a central policy layer for estate-wide rules.
How do you stop AI agents surfacing overshared SharePoint files?
Use SharePoint Advanced Management in order: DAG reports, restricted content discovery, restricted access control, then site access reviews.
How do you limit who can share Copilot Studio agents?
Use Managed Environments sharing rules to block security-group sharing, cap viewers and restrict Editor rights for Copilot Studio agents, and know the gaps.
Where are Copilot Studio conversation transcripts stored and for how long?
Copilot Studio transcripts live in the Dataverse ConversationTranscript table and are deleted after 30 days by default. How to change it and who can read them.
How do you control which ChatGPT Enterprise users get agent features?
Use ChatGPT Enterprise custom roles assigned to SCIM-synced groups to roll out agent features, apps and Codex to some users only. How permissions combine.
How long does LangSmith keep agent traces, and is it enough for audit?
LangSmith keeps traces 14 days (base) or up to 180 days (extended). What upgrades a trace, how to export, and where audit gaps remain.
Put the framework into practice