A fast security gateway for AI agent tool calls
When an AI agent is connected to a database, a local shell, or an API, a model that can behave unpredictably gets access to real actions. Common safety approaches either inspect the model’s text and hope nothing dangerous slipped through, or add another large language model as a judge, which can create a large delay.
Trajeckt puts a sealed gateway at the execution layer instead of trying to clean up the prompt or catch suspicious words. If a tool call or execution path does not match the preset spec exactly, it is dropped.
The claimed added delay is about 1.6 milliseconds, much lower than a 500 millisecond safety check. The gateway is invisible to the model, so the agent cannot escape it through prompt injection or jailbreak tricks because the decision is enforced outside the model.
Key points
- AI agents become riskier when they can use databases, local shells, or APIs.
- Model-as-a-judge safety checks can add cost and delay.
- Trajeckt checks real tool calls at the execution layer.
- Calls that do not match the spec are rejected immediately.
- The claimed added delay is about 1.6 milliseconds.
Quick term guide
- large language model
- The type of AI behind ChatGPT or Claude — trained on huge amounts of text to read, write, and code.
- execution layer
- The part of a system where requested actions actually run.
- safety check
- A step that checks for problems before the tool continues.
- prompt injection
- A trick where hidden instructions in text make an AI do something the user did not ask for.
- jailbreak
- An attempt to bypass an AI system’s safety limits.
- reliability
- How consistently a tool works without failing or behaving unexpectedly.
- production
- The live version of a service that real users use.
- tool calls
- Times when an AI system uses another function, such as search or file access.