#099Guide3 min

Assess risks of AI agents

Review agent instructions, tools and components before giving them access to your systems.

Hackaiz team · updated 21/11/2026

Illustration: a professional files evidence folders.
Illustration: a professional files evidence folders. Illustrative photo, not a screenshot.

01Describe what the agent can actually do

An AI agent combines model-generated decisions with tools and instructions. Its practical risk depends on the authority those tools carry: reading a local file is different from changing a production database or sending a deployment command. Before checking a component, make your own inventory of its inputs, credentials, permitted destinations and actions. Record where a human must authorize a consequential step. This inventory is preparation by your team, not an automatically discovered permissions map.

Consider a support agent that reads incoming messages and can call an account-management tool. A message may contain instructions intended to divert the agent from its assigned task. That is an example of prompt injection: untrusted content tries to become an instruction. If the same agent can export records without a separate authorization check, the tool’s authority can turn a misleading instruction into a data exposure. The example explains a risk, not a claim that Hackaiz has observed it in your project.

02Inspect components before installation

Sign in and open Check before installing at /check. Paste the Skill instructions from SKILL.md or an MCP configuration into the text area, or select Or choose files to provide component files. MCP means Model Context Protocol, a way for an assistant to use external tools. The supported component workflow also accepts material such as package.json, a GitHub workflow or a Dockerfile. Use the component name field to make the result recognizable, then select Check.

This check reads the submitted material without executing it. Do not paste real credentials into instructions or configurations; secret files are excluded during file selection. Review the list labelled Excluded files, not checked and any incomplete-coverage message before interpreting the result. Every new check uses credits, including a repeat with identical content. File contents are not retained by this workflow; only the result summary is kept in history, so preserve your own reviewed version separately.

03Read findings in terms of consequences

Look for the concrete behavior behind the warning: a hidden command, a read of secret material, a destination receiving data, or an instruction attempting to redirect the assistant. In application findings, the customer priorities are Danger, To fix and To watch. The component check also has its own overall verdict, including Incomplete when the material could not be fully assessed. Do not translate an incomplete result into approval to install or assume that a lower-priority finding cannot matter in your deployment.

The AI reference mappings distinguish several concerns. Tool-related rule patterns can map to LLM06 Excessive Agency; memory-related patterns can map to poisoning risks, where untrusted material contaminates information reused later. The code also includes a signal for selected AI call paths with neither a recognizable token limit nor rate limiting. These signals guide a review of authority, stored context and consumption. They do not prove that all model behaviors, runtime permissions or possible conversations have been tested.

04Reduce authority and verify the changed version

Ask the component owner to justify each capability. As operational precautions, prefer a read-only credential where writing is unnecessary, restrict allowed destinations, separate customer data and require human approval for destructive actions. Test changes in a controlled environment with non-sensitive data. These are recommendations for your own agent setup, not controls that the static component check automatically installs or enforces.

Recheck a changed component and review the evidence before widening access. The results screen can flag new capabilities or new rule signals since the previous check; investigate what changed rather than treating an older result as permanent approval. If your assessment also includes a website or API, Hackaiz requires domain ownership proof before analysis, through the supplied DNS TXT record, HTTPS verification file or your project’s AI creating that file. A component inspection does not bypass that authorization boundary.

Frequently asked questions

Was this article helpful?

Related articles