The INTRUSIVE.AI platform

INTRUSIVE.AI uses simulations and synthetic environments to help you understand how an AI agent might respond to various scenarios in production. This is especially helpful for exposing indirect prompt injection risks that are unique to your specific agent architecture but it can also be helpful for evaluating basic operational aspects of agent performance like task consistency and reliability.

Three components, one workflow

Component What it does Start here
Operator Sandbox Runs campaigns, controls Attack Harness tools, enforces limits, and retains results. Operator
Attack Harness Uses AI to plan experiments, submit them through Operator, and adapt to feedback. Attack Harness
Interceptor Sandbox Runs your agent with simulated dependencies, controlled data, and recorded observations. Interceptor

Operator manages the Attack Harness container. Interceptor manages the target agent's environment. The component conducting the investigation and the agent being investigated do not share unrestricted access to one another or the network.

Your own AI harness can use the INTRUSIVE.AI skills to help configure the platform and prepare reports. That assistant is distinct from Attack Harness, which performs the campaign's adaptive testing.

Evaluate security and functional effectiveness

Attack simulations explore how adversarial prompts could affect the agent in production, including indirect prompt injection through retrieved content, tool responses, or other inputs specific to its architecture. Operational simulations explore whether the agent performs its intended tasks effectively: producing correct answers, completing requested work, and behaving consistently across relevant cases.

Both use your chosen objectives and the evidence available from the synthetic environment. For example, a customer-support investigation might examine whether untrusted content redirects the agent, or whether ordinary order-status requests receive accurate, consistent answers. Choose either focus or include both with separate evidence criteria. These are investigation choices within the campaign workflow, not separate product modes.

Results show what happened under the tested conditions. Use them to assess how the agent might behave in production, accounting for differences in data, dependencies, configuration, and workload between the synthetic environment and production.

From an agent to a report

  1. Bring the agent. Package its code and dependencies in a container image.
  2. Describe its world. An Interceptor environment blueprint defines how to invoke it and the services, files, identities, and data it will encounter.
  3. Connect Operator. A private target profile sets permitted operations and the destination. Generated capability exports describe the available interface.
  4. Describe the investigation. A scenario bundle contains your objectives, optional scenarios, evidence expectations, and requested limits.
  5. Run a campaign. Operator starts Attack Harness, delivers permitted requests, and records results while Interceptor observes the target.
  6. Review the evidence. Inspect the retained run and generate a report that separates observations, interpretations, and unanswered questions.

Interceptor is intended to supply the simulated services an agent needs inside its synthetic environment. Its managed PostgreSQL/pgvector service, for example, lets a database-backed agent run without a separately operated database. Supported service types still need configuration from the application's actual dependencies; an undeclared dependency does not automatically become a working simulation.

Choose your starting point

Choose documentation for your release

Use the documentation's product selector, then choose the version you installed. Product versions are independent; selecting Operator does not select compatible Attack Harness or Interceptor releases. Archived documentation remains selectable. If a page is missing in a selected release, navigation explains the fallback to that release's overview.

Start at the official releases site for stable packages and images, and check integration contracts before running the components together.