Why AI agents are a risk
An AI agent decides what to do based on the content it reads, and an attacker can write some of that content. The agent then acts with the IAM role and the tools you gave it. For example, an AI agent for customer support answers messages and can look up orders. An attacker sends a message that tells the agent to paste another customer’s orders into the reply. The agent treats the message as an instruction and follows it. Its IAM role and tools allow the lookup, so the attacker gets the orders. To limit the damage, you need to know which agents you have and which roles and tools they use.How it works
Plerion reads your agents’ configuration through your AWS, GCP and Azure integrations. You do not install anything or add code to your agents. Agent security covers agents on Amazon Bedrock, on Google Cloud and on Microsoft Foundry. Agents you host yourself, outside these services, are not covered. On AWS, Amazon Bedrock AgentCore is a service for building and running AI agents. Agent security covers three kinds of AWS agent:- AgentCore runtimes: you supply the agent code or a container image.
- AgentCore harnesses: you define a model, tools, and instructions.
- Amazon Bedrock Agents Classic agents: built with the earlier Amazon Bedrock Agents service.
- Vertex AI Agent Engine agents: deployed from your own code or container image.
- Gemini Enterprise agents.
- Dialogflow CX agents.
What Plerion shows about your agents
OpenAgent security in the left navigation to see your agents in every account and region. For these agents, Plerion shows:
- For AWS and GCP agents, their risk score
- For AWS agents, the IAM role they run as
- For AgentCore runtimes and harnesses, whether they run in a VPC and how their callers are authorized
- For AgentCore runtimes, the tools their code defines, on the Tools tab
- For AgentCore runtimes, gateways and Amazon Bedrock Agents Classic agents, what each one can access and what can access it, on the Access tab, as a graph or a list with IAM and resource policies applied
- For AgentCore harnesses, the gateways they call tools through, their AgentCore memory, and the credential providers they use to sign in to other services
- For Vertex AI Agent Engine agents, the vulnerabilities, secrets and software bill of materials (SBOM) in their code or image, and the tools their code defines, when GCP workload scanning covers them
- For Dialogflow CX agents, the tools, webhooks and playbook code actions they are set up with, on the Tools tab, and a one-sentence summary of what the agent is for
- For Microsoft Foundry agents, their tools on the Tools tab and a one-sentence summary of what the agent is for. A prompt agent’s tools are the ones it is set up with, and a hosted agent’s are the tools its code defines.
- A gateway anyone can call
- A gateway with no policy engine
- A policy engine that logs decisions without enforcing them
- A Cedar policy that allows any caller
- A runtime, gateway, or memory that an account outside your AWS organization can access
- Secrets left in an agent’s environment variables or configuration
- A runtime or harness that runs outside a VPC
- A harness that calls a remote tool server over plain HTTP
- Vulnerabilities and secrets in a runtime’s container image
- An Amazon Bedrock Agents Classic agent with no guardrail
- A region with model invocation logging turned off
- Vulnerabilities and secrets in a Vertex AI Agent Engine agent’s code or container image
- A Vertex AI Agent Engine agent with a secret in a plaintext environment variable, or with no Private Service Connect interface
- A Gemini Enterprise agent shared with all users
- A Dialogflow CX agent with no security settings attached, or with Cloud Logging turned off