#137Case study3 min

Can your AI agent become a security risk?

Separate model instructions, tool permissions and component checks before trusting an agent.

Hackaiz team · updated 21/11/2026

Illustration: a creator films a short video on a phone.
Illustration: a creator films a short video on a phone. Illustrative photo, not a screenshot.

01The risk follows the tools, not just the model

An AI agent combines a model with instructions and tools that can act on its behalf. A writing assistant and an assistant allowed to read files, call external services or prepare code changes do not have the same exposure. Before connecting another tool, list what the agent can read, where it can send information and which actions require your approval. Treat this as a practical review of permissions, not a judgement based on the model name.

For example, a document supplied to an assistant may contain instructions to disclose private information. That is an instruction-injection risk: content that should be treated as data tries to become an order. Excessive permissions make the consequences worse. These are illustrative risk scenarios, not accounts of incidents at Hackaiz customers. A well-written instruction alone does not establish that every tool call is safe.

02Check a component before giving it access

Hackaiz provides a security plugin for Cursor that can check a Skill, an MCP server, a dependency, a CI workflow or a Dockerfile. A Skill supplies reusable instructions; an MCP server connects an assistant to tools; a CI workflow describes automated repository tasks. These components can introduce behaviour beyond the conversation itself. Review the component before installing or running it rather than assuming that an assistant recommendation is a security review.

With the plugin installed, ask Cursor “is this safe?” about the component. The documented requirements include Node 22 or later, a Hackaiz account and an API key in the HACKAIZ_API_KEY environment variable. Never paste that key into the conversation. The agent requests consent before each check, and every check consumes account credits. Use the documented installation instructions rather than sharing credentials in a prompt.

03Understand what a component verdict actually covers

The component check is static: Hackaiz reads the submitted content without executing it. Its verdicts are Safe, Warning, Danger and Incomplete, accompanied by available file names, line numbers, rule identifiers and recommended actions. These are the component-check verdicts, not the severity groups used in project security reports. Keep the exact file and rule references when discussing a result with a developer.

A Safe verdict means the static check found no known risk pattern in the inspected content. It does not reveal runtime behaviour or the full code of installed dependencies, and it is not permission to grant unrestricted access. If a result is Incomplete, establish what was not inspected before deciding whether to proceed. If a risky command is identified, examine its actual purpose and permissions rather than simply asking the model to reassure you.

04Review the system around the agent as well

A component check does not replace an authorized review of the application that hosts an agent. Use Connect for the system you control, then examine the findings for the connected project. Websites and APIs require domain ownership proof before analysis: a DNS TXT record named _hackaiz, the HTTPS file /.well-known/hackaiz-verification.txt, or an instruction to the project’s AI to create that file. Merely knowing an endpoint is not authorization.

Project reports group severity as Danger, To fix and To watch. Review the evidence and scope of each finding, especially whether it concerns instructions, data exposure or permissions. Code fixes, when available, remain proposals: a human reviews and approves them before delivery as draft GitHub pull requests, and Hackaiz never merges them automatically. The Cursor component checker itself never modifies, publishes or merges code. Keep installation approval separate from approval of a code change.

Frequently asked questions

Was this article helpful?

Related articles