Agentic Tool Use

How Tool Use Works

The application describes each available tool to the model with a name, a purpose and the inputs it accepts. When the model decides a tool would help, it returns a structured request instead of a final answer. The application runs the tool, passes the result back and the model continues. An agent may repeat this loop many times before it finishes.

Why Tool Use Changed What Models Can Do

Without tools, a model can only reason about the information it was given. With tools, it can go and check. In code security, that means a model can open the caller of a function, search for where input is validated or read a configuration file, instead of guessing from a short snippet.

Tool use also makes cheaper models more useful. A smaller model that can look up the evidence it needs can often reach the right answer on tasks that once needed a much larger model.

Security Risks of Tool Use

• Excessive agency. Tools with more power or permissions than the task needs.
• Prompt injection. Untrusted content in a file, page or ticket that tries to steer the agent into misusing its tools.
• Unvalidated output. Tool results or model output passed to another system without checks.
• Unclear accountability. Actions taken by an agent with no record of why.

Treat every tool as code that runs with real permissions, and scope it accordingly.

How Nullify Uses Tool Use

Nullify's triage agents investigate findings with the repository and asset graph in reach. In one rebuilt screening stage, a cheaper model that could read, grep and glob the repository itself used 97.5% fewer input tokens and ran in 18.4 seconds instead of 592.9, while getting the answer right where the old design had deferred it. See how Nullify investigates findings.