82% of orgs found shadow agents (CSA, 2026)

You know AI is running in
your cloud.

You don't know what it can touch.

Oscar scans your live cloud environment and surfaces every AI agent, its IAM scope, production system access, and blast radius. In days, not weeks. No credentials leave your perimeter. No changes made to your environment.

Get the Assessment See how it works
Read-only scan Findings in 5-10 business days No credentials stored centrally Setup under 30 minutes
82%
Of orgs found AI agents they didn't know about (CSA, 2026)
+218%
Typical 90-day token spend growth from untracked agent deployments
Days
Not 8-16 weeks. How long the Assessment takes vs. a consulting firm
5
Evidence-backed deliverables from live infrastructure, not interviews
The Problem

Agents get deployed. Nobody tracks them.

When developers adopt Claude Code, Cursor, GitHub Copilot, or custom LangChain workflows, those tools leave a real footprint. IAM roles get created. Lambda functions get deployed. MCP server connections get established. CI/CD pipelines get write access. Nobody files a ticket.

Shadow agents with no owner

Engineers deploy agents using AI tools without a formal provisioning process. Six months later, the agent is still running, still touching production systems, and nobody knows who built it.

Permissions far broader than needed

Agents get deployed with PowerUser or AdministratorAccess because it was faster at the time. The permissions are never right-sized. The blast radius grows quietly alongside your environment.

Untracked token spend

Token costs accumulate across agents nobody budgeted for. Finance sees a growing API bill. Nobody can attribute it to a team, a workflow, or a business purpose.

Direct access to production systems

Agents reading PII S3 buckets, writing to payment databases, or deploying to production ECS clusters with no approval gate and no documentation of who authorized the access.

No compliance evidence

SOC 2, PCI DSS, SR 11-7, and the EU AI Act all require provable governance of AI systems in production. If you cannot name the agents, you cannot govern them. If you cannot govern them, you cannot attest to them.

Manual discovery takes months

A Big Four consulting engagement to build this picture takes 8 to 16 weeks and costs $75K to $300K. By the time the report lands, the environment has already changed. And none of it updates automatically.

The Governance Gap

Five questions you cannot currently answer.

These are not hypothetical questions from a future auditor. They are questions your CISO, your board, and your cyber insurer are already asking.

1
How many AI agents are running in your cloud right now?
2
Which ones have access to production databases or PCI-scoped systems?
3
What is the blast radius if one of them gets a bad instruction or a compromised credential?
4
Who owns each agent, and who would know if it started behaving differently?
5
How much are they costing you, and is any of that spend in your budget?
What Oscar Actually Finds

This is not a theoretical risk.

A recent assessment of a mid-market financial services firm found 34 AI agents running across 11 AWS accounts. Here is what was inside.

Real findings. Anonymized environment.

19 shadow agents. 6 critical findings. $31,400/month in unbudgeted spend. One agent with AdministratorAccess across all 11 accounts -- running for 67 days with no documented owner.

Oscar identified 6 Lambda functions in the payments-prod account making direct API calls to core banking services. These functions matched the signature of AI agent workflows -- no IaC lineage, broad IAM scope, irregular call patterns in CloudTrail -- but did not appear in the approved AI tools inventory. A cost-optimization agent with EC2 termination rights across 3 accounts had no environment tag filter. CloudTrail confirmed it terminated 4 production instances the same night a 3-hour outage was logged. Causal connection flagged for investigation.

34
Total agents discovered across 11 accounts
19
Shadow agents with no inventory record
73%
Of monthly token spend was unbudgeted
6
Critical findings requiring immediate action
How It Works

Read-only discovery. Evidence-backed findings.
Human-approved at every gate.

Oscar builds the Context Graph from your live environment automatically. No tagging required. No credentials leave your perimeter. No changes made to your infrastructure.

1

Connect with read-only credentials

Oscar connects to your AWS accounts using existing IAM read-only credentials. Setup takes under 30 minutes. Nothing is stored centrally. No new permissions are required beyond what Oscar needs to observe your environment.

2

Oscar scans IAM, CloudTrail, Lambda, MCP, and CI/CD simultaneously

Agent discovery requires correlating data across IAM role patterns, CloudTrail behavioral signatures, Lambda deployment history, MCP server process detection, and CI/CD pipeline inspection at the same time. Oscar does this automatically. A consulting team doing this manually takes 8 to 16 weeks. Oscar completes the scan in hours.

3

Findings reviewed and confirmed by human operators

Every finding is reviewed against your stated AI governance policies and tool inventory before it enters the report. Oscar surfaces blockers and gaps as named items requiring human resolution. Nothing advances on assumed evidence.

4

You receive the complete assessment report

Full agent inventory, blast radius map for each high-risk agent, token cost attribution by agent and team, governance gap analysis against your compliance framework, and a prioritized remediation plan with ownership mapped to the right teams. Delivered in 5 to 10 business days.

What You Get

Five deliverables. All from live infrastructure.

Every deliverable is grounded in what Oscar actually observed in your environment. Not assembled from interviews. Not inferred from documentation. Not stale by the time it reaches you.

Complete Agent Inventory

Every AI agent and automated workflow running across your cloud accounts, verified from live infrastructure. Includes agent type, deployment method, IAM scope, account location, and documentation status.

Blast Radius Map

For each high-risk agent: what production systems are reachable from its current IAM scope, whether it can modify IAM, what accounts it can affect, and how reversible its actions are. What it could affect, not just what it does.

Token Cost Attribution

Monthly token spend broken down by agent and team, including unbudgeted shadow spend. Correlated from CloudWatch logs, Lambda invocation records, and API gateway telemetry. Budget gaps called out explicitly.

Governance Gap Report

Where your current agent posture diverges from SOC 2, PCI DSS, SR 11-7, or your own stated AI governance policies. Each gap includes affected resources, the applicable compliance control, and the specific remediation step.

Prioritized Remediation Plan

Specific steps, with ownership mapped to the right teams, and the IaC or CLI command to execute each one. Not a consultant's deck of recommendations. Actionable steps your team can execute immediately.

Add-on

Ongoing Monitoring

The assessment is your starting point. Monitoring keeps the inventory current. New agent detected, alert fires. Permission scope change, alert fires. Spend crosses budget threshold, alert fires. Stay current without running another engagement.

Why OpsCanvas

The hard part is discovery, not analysis.

Identifying AI agent footprint requires correlating 5 data sources at once.

Manual discovery requires correlating data across IAM configurations, CI/CD pipelines, CloudTrail logs, deployment history, and running resource inventories simultaneously. Doing this by hand takes weeks and produces a snapshot that is already partially stale the moment it lands.

Most organizations attempting this today are either working with consultants who charge $75K to $300K for 8 to 16 weeks of work, or they are not doing it at all.

OpsCanvas delivers this in days because the Context Graph does the discovery automatically. The data that would take a consulting team weeks to assemble is available to Oscar in hours.

No manual inventory or tagging required
Multi-account, multi-cloud coverage in a single scan
Evidence artifacts, not interview summaries
Human-approved at every decision gate
Findings in 5 to 10 business days
Capability
OpsCanvas
Discovery method
Live infrastructure scan
Time to findings
5-10 business days
Manual inventory required
No
Tagging required
No
Token cost attribution
Per agent, per team
Blast radius analysis
Included
Continuous monitoring
Available as add-on
Evidence artifacts
CloudTrail-backed
Stay Current

What happens next week when your engineers
deploy three more?

An assessment tells you where you stand today. Agents get deployed every week. OpsCanvas Monitoring keeps the inventory current, automatically.

1

Assessment complete

You have a verified inventory, blast radius map, and remediation plan from the initial engagement.

2

New agent deployed

An engineer ships a new LangChain workflow on a Tuesday. No ticket filed. No inventory updated.

3

Oscar detects it

Monitoring picks up the new IAM role, Lambda function, and CloudTrail signature. Alert fires within hours.

4

Inventory stays current

The agent is added to the live inventory. Scope and blast radius are assessed. Owner is routed for confirmation.

Common Questions

Frequently asked.

Oscar scans all AWS accounts you connect using read-only credentials. Multi-account environments with AWS Organizations are supported. You can scope the scan to specific accounts or run it across your entire organization. Region coverage is configurable.

The AI Agent Inventory Assessment currently focuses on AWS, where the majority of enterprise AI agent footprint lives today. Multi-cloud coverage is on the roadmap. If your environment is primarily Azure or GCP, talk to us -- we can discuss what is feasible for your situation.

Oscar uses a combination of signals: IAM role naming patterns consistent with AI tool provisioning, environment variable inspection for LLM API keys and model references, CloudTrail behavioral signatures showing irregular invocation patterns, MCP server process detection, and CI/CD pipeline inspection for AI-specific deployment signatures. Oscar does not rely on any single signal. An agent flagged as high-confidence has been identified via multiple corroborating evidence sources.

Blast radius reflects what could be affected by a misconfiguration, prompt injection, or compromised credential -- not what the agent is designed to do. For each high-risk agent, Oscar maps the production systems reachable from its current IAM scope, whether it can modify IAM itself, which accounts it can reach, and how reversible its actions are. This gives you the true risk exposure, not just the intended function.

The assessment is a scoped engagement that gives you a complete picture as of a specific date. The monitoring add-on keeps that picture current on an ongoing basis. It detects new agent deployments, flags IAM scope changes, tracks token spend against budgets, and alerts on governance drift. The two are designed to work together: start with the assessment to establish a baseline, add monitoring to ensure you never fall back into the dark.

The governance gap report maps findings against SOC 2, PCI DSS, SR 11-7 (Model Risk Management), and CCPA by default. If your organization operates under HIPAA, the EU AI Act, or sector-specific frameworks, let us know during scoping and we will configure the gap analysis accordingly.

Get Started

Start with the assessment.
Stay current with monitoring.

Oscar scans your live environment with read-only credentials. No changes made. No credentials stored centrally. Setup under 30 minutes. Findings in 5 to 10 business days.

Read-only access  |  No changes to your environment  |  Setup under 30 minutes