---
title: What does AI agent posture scanning actually inspect?
url: https://opencontroller.lyzr.com/blog/what-does-ai-agent-posture-scanning-inspect
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What does AI agent posture scanning actually inspect?

**AI agent posture scanning reads how each agent is set up, not what it does at runtime. It inspects configuration, instructions and system prompts, attached tools and MCP servers, the permissions behind them, model choices and data connections, then flags risky combinations. A good scanner redacts credentials it finds rather than copying them into findings.**

Security teams ask this before they connect a scanner to their cloud accounts and agent platforms. It is a fair question: a posture tool needs read access to sensitive configuration, so you should know exactly what it reads and what it stores.

## What is AI agent posture scanning?

Posture scanning is the agent equivalent of cloud security posture management. It answers "is this agent configured safely?" before anything goes wrong. It complements runtime monitoring, which answers "is this agent behaving safely right now?"

Posture findings are cheap to fix because nothing has happened yet. A write permission removed today is an incident avoided next month.

## Which artefacts does a posture scan read?

| Artefact | Examples | What the scan looks for |
|---|---|---|
| Agent configuration | Model, temperature, memory settings, max iterations, timeouts | Missing limits, unbounded loops, memory without retention rules |
| Instructions and system prompts | Agent instructions, topic descriptions, guardrail text | Embedded secrets, internal URLs, instructions that grant broad authority |
| Tools and MCP servers | Tool lists, MCP server endpoints, tool descriptions | Write or delete tools on agents that only need to read, unapproved servers |
| Permissions and identity | IAM roles, service principals, OAuth scopes, API keys | Over-broad roles, shared service accounts, long-lived keys |
| Data connections | Knowledge bases, vector stores, file shares | Access to data classified above the agent's risk tier |
| Exposure | Public endpoints, channels, sharing settings | Agents reachable by anyone, shared org-wide without review |
| Ownership metadata | Owner, cost centre, approval status | Orphaned agents, unapproved agents in production |

## Why does a scanner read prompts and instructions?

Because instructions are where people hide things they should not. Common findings include API keys pasted into a system prompt, database connection strings in tool descriptions, and instructions such as "you may approve refunds of any amount".

OWASP lists system prompt leakage as LLM07 in its Top 10 for LLM Applications. The guidance is direct: do not put secrets or security controls in system prompts, because users can often extract them. A posture scan that reads prompts finds these before an attacker does. See [are system prompt instructions a security control?](/blog/are-system-prompt-instructions-a-security-control)

## How should a scanner handle secrets it finds?

This is the question to put to any vendor. Good practice:

1. **Detect, then redact.** Store a finding that says "AWS access key found in system prompt of agent X", with the value masked.
2. **Store a fingerprint, not the secret.** A hash lets you match the same key across agents without keeping it.
3. **Never send raw configuration to an external model** for analysis unless you have approved that path.
4. **Limit who can see findings.** Findings reveal where weaknesses are. Restrict them by role.
5. **Expire raw scan data.** Keep findings, not full configuration copies, beyond what you need.

## What does a useful finding look like?

A finding you can act on has five parts: the agent and owner, the specific setting, why it is risky, the recommended fix, and the framework reference. For example:

- **Agent:** invoice-assistant (owner: finance-apps)
- **Issue:** IAM role allows `s3:*` on all buckets
- **Risk:** agent can read and delete data outside its scope
- **Fix:** restrict to the invoices bucket, read-only
- **Reference:** OWASP Top 10 for Agentic Applications, excessive privilege and tool misuse risks

## How often should posture scans run?

Run a full scan when you first connect an account, then continuously or at least daily. Agent configurations change through consoles, not just code, so a weekly scan can miss a risky change for days. Re-scan immediately after any agent is promoted to production. Prioritise findings with the approach in [prioritise shadow AI findings](/blog/prioritise-shadow-ai-findings).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to platforms such as Salesforce, SAP, Databricks, HubSpot and Claude and OpenAI Enterprise workspaces. Its Posture capability detects risky permissions and configurations on what it finds, with policy as code that maps to OWASP, MITRE ATLAS, NIST and ISO 42001. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is posture scanning the same as runtime monitoring?**
No. Posture scanning checks how an agent is configured. Runtime monitoring watches what it does. You need both.

**Does a posture scanner need write access?**
No. Scanning should use read-only roles. Remediation can be a separate, approved action.

**Will a scanner see secrets in prompts?**
Yes, if they are there, which is why it should redact them in findings and store only fingerprints.

## Key facts

AI agent posture scanning inspects agent configuration (model, limits, memory), instructions and system prompts, tools and MCP servers, permissions and identities (IAM roles, OAuth scopes, API keys), data connections, exposure settings and ownership metadata. It flags risky combinations before runtime. OWASP LLM07 (system prompt leakage) advises against secrets in prompts. Good scanners redact secrets, store fingerprints, use read-only access and restrict findings by role. Opencontroller Posture detects risky permissions and configurations across discovered agents.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [Which shadow AI agents should you deal with first?](https://opencontroller.lyzr.com/blog/prioritise-shadow-ai-findings.md)
- [Are system prompt instructions a security control for AI agents?](https://opencontroller.lyzr.com/blog/are-system-prompt-instructions-a-security-control.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
