Start typing to search this publication.
techcirkle
Open menu

Subscribe to techcirkle

Get new posts delivered straight to your inbox.

Your AI Stack Is Now Cloud Attack Surface: What CNAPPs See and Miss

Inference endpoints, agent credentials, vector databases and model API keys add risk. What AI-SPM in CNAPP tools catches and what you must design.

techcirkle
5 min read
Cover image for Your AI Stack Is Now Cloud Attack Surface: What CNAPPs See and Miss
Padlock over a cloud network representing security for AI and LLM workloads
Padlock over a cloud network representing security for AI and LLM workloads

Most cloud security programmes were designed around a familiar set of assets: VMs, containers, databases, storage, identities. Over the last two years, product teams have added a new set, often quickly and sometimes outside the standard landing zone: model inference endpoints, LLM-powered agents, vector databases and a growing collection of third-party model API keys.

Infrastructure-wise these are ordinary services, identities and data stores. What makes them risky is how they are combined, how fast they ship, and the fact that an AI agent turns natural-language input into privileged actions.

This post walks through the four main areas of exposure, what AI security posture management (AI-SPM) features in cloud-native application protection platforms can detect, and which controls only exist if you build them in. For the wider platform landscape, see our full guide to CNAPP tools.

1. Inference endpoints

The risk. Self-hosted models and managed inference services are frequently stood up for a prototype, then promoted to production with prototype settings: public exposure, no authentication or a shared static key, broad read access to training or reference data, GPU nodes in an account that has not been onboarded to central logging.

What a CNAPP can catch. This is familiar ground. Posture checks find publicly reachable endpoints, missing authentication on managed services and disabled logging. Workload scanning finds vulnerable packages in model-serving images, and inference frameworks pull in large dependency trees. AI-SPM features add discovery, identifying which services, notebooks and managed AI resources exist so shadow deployments show up in inventory.

What you must design. Put inference behind your internal gateway with authentication, rate limits and per-tenant quotas. Decide upfront what data the serving identity may read, and keep model artefacts in versioned, access-controlled storage.

2. Agent credentials and prompt injection

The risk. An agent that can read tickets, query a database, call internal APIs and trigger a deployment is one of the most privileged identities in your estate. Prompt injection, meaning malicious instructions hidden in content the agent reads such as an email, a web page, a document or a ticket comment, can redirect it. At that point the agent's permissions are effectively the attacker's permissions.

What a CNAPP can catch. Entitlement analysis can flag that the identity behind an agent holds wildcard or unused permissions, can assume roles across accounts, or can reach sensitive data stores. Runtime detection may notice anomalous API usage from that identity. Some AI-SPM features inventory agent tool permissions.

What a CNAPP cannot catch. It has no view of your prompt, your tool definitions or the logic deciding when a tool is called. It cannot tell a legitimate instruction from an injected one. That is application design.

What you must design:

  • Scoped credentials per tool. Each tool gets its own narrowly permissioned, short-lived credential, not one shared agent role.

  • Human approval gates for irreversible actions. Deletions, deployments, payments and external messages require explicit confirmation outside the model's control.

  • Treat retrieved and user content as untrusted. Keep system instructions separate from data, and validate tool arguments against schemas and allow-lists before execution.

A tool policy can be made explicit in configuration. The example below is illustrative:

# Illustrative agent tool policy
tools:
  search_tickets:
    credential: role/agent-tickets-readonly
    approval: none
  query_orders_db:
    credential: role/agent-orders-read-scoped
    row_filter: requesting_user_tenant
    approval: none
  trigger_deploy:
    credential: role/agent-deploy-staging-only
    approval: human_required
    allowed_environments: [staging]

When we deliver agentic workflow development, policies like this sit alongside the code and get reviewed like any other change.

Engineer reviewing AI agent permissions and logs on a monitor
Engineer reviewing AI agent permissions and logs on a monitor

3. Vector databases and retrieval pipelines

The risk. Retrieval-augmented generation copies documents into a new store as chunks and embeddings. That store frequently ends up with weaker access control than the source systems. A support-bot index built from internal wikis, contracts and tickets now holds data that used to be permissioned per team or per customer.

What a CNAPP can catch. DSPM can discover managed vector stores and databases, flag public exposure and missing encryption, and in some cases classify sensitive content. Coverage for newer vector database products and for embeddings specifically varies between platforms, so test it rather than assume it.

What you must design. Permission-aware retrieval. Carry source-system access metadata into each chunk, and filter results by the requesting user's permissions before anything reaches the model. A scanner can tell you the store is sensitive. It cannot tell you that user A just retrieved a paragraph only user B should see.

4. Third-party model API keys

The risk. Keys for external model providers end up in environment variables, notebooks, CI secrets, front-end bundles and chat messages. Each is also an outbound data path: whatever you send in a prompt leaves your boundary.

What a CNAPP can catch. Secret scanning in repositories, images and IaC catches many committed keys. Posture and network views can surface unexpected egress.

What you must design. Route model calls through a single internal proxy that holds provider keys, applies redaction policies, enforces per-team budgets and records usage.

The control that ties it together: logging

Across all four areas, the most valuable single control is complete, structured logging of prompts, retrieved context references, tool calls, arguments, approvals and outcomes, tied to the user and agent identity. Without it, incident response for an AI system is guesswork. A CNAPP can ingest cloud audit logs, but it cannot reconstruct why an agent did what it did unless your application recorded it.

A practical split of responsibilities

Area

CNAPP / AI-SPM can detect

Must be built in

Inference endpoints

Public exposure, missing auth, vulnerable images, shadow deployments

Gateway, quotas, scoped data access

Agents

Over-privileged identities, anomalous API use

Per-tool credentials, approval gates, input handling

Vector stores

Exposure, encryption, some classification

Permission-aware retrieval

Model API keys

Committed secrets, unusual egress

Central proxy, redaction, rotation

If you are integrating models into production systems and want these controls designed in from the first sprint, our LLM integration team works this way by default.

Frequently Asked Questions

What is AI security posture management?

AI-SPM is a set of CNAPP features that discover AI services, model endpoints, related data stores and agent permissions across a cloud estate, then flag misconfigurations such as public exposure, missing authentication or excessive access.

Can a CNAPP detect prompt injection?

Not directly. It can flag that an agent's identity is over-privileged or behaving unusually, but it cannot see prompts, tool logic or the difference between legitimate and injected instructions. Defences must be built into the application.

Why are AI agents considered high-risk identities?

Agents often combine read access to many systems with the ability to take actions, and their behaviour can be influenced by untrusted content. If manipulated, they can exercise every permission they hold.

How should vector databases be secured?

Apply standard exposure and encryption controls, then enforce permission-aware retrieval so each query only returns chunks the requesting user is entitled to see in the source system.

What is the safest way to manage third-party model API keys?

Keep keys out of application code by routing calls through an internal proxy that stores them, applies redaction and budgets, logs usage and supports regular rotation with separate keys per environment.

What should be logged for LLM and agent systems?

Prompts, references to retrieved context, tool calls with arguments, approval decisions and outcomes, all linked to the user and agent identity, so incidents can be investigated and behaviour audited.

Platform categories, architectures and evaluation guidance are in the full guide: CNAPP Tools in 2026.

Subscribe to techcirkle