Intrusive.ai
Sandboxed attack simulation for testing agentic AI systems for consequential failures
- Platforms
- API, self-hosted
- Listing updated
- today
Similar tools
About Intrusive.ai
Intrusive.ai is an AI tool for cybersecurity. It is available on API and self-hosted.
The site describes a method for exercising AI agents against adversarial input inside synthetic, restorable environments. The agent's tools, identities, services, and data run in a sealed world where outbound network access is replaced by synthetic endpoints, so an attempted action such as a file read or an HTTP egress can be observed and recorded without touching production systems.
The platform is split into three parts: an Operator Sandbox that contains the offensive testing process, an Attack Harness that adapts attempts based on what it observes, and an Interceptor Sandbox that places a real agent application inside a synthetic operating world. Model calls from the target pass through a host-side connector, and recorded observations and logical snapshots let testers replay and compare runs.
The product is aimed at AI security testers, red teams, and teams building agent applications who need to verify whether guardrails and approvals actually stop an action, not just change the agent's reply. The company states that its operator, harness, and interceptor components are open source under AGPL-3.0-only.
What it does
- Sealed synthetic environment
- Runs the target agent in a network-isolated world with synthetic services, data, and attacker endpoints.
- Causal failure traces
- Records the path from adversarial input through state changes and tool calls to the final external effect.
- Restorable state snapshots
- Lets testers restore declared state, branch runs, and compare repeated trials in fresh sessions.
- Scenario coverage library
- Includes scenarios such as indirect prompt injection, capability escalation, credential misuse, and data exfiltration.
- Host-mediated model access
- Routes supported model calls through a host-side connector with explicit credentials and limits.
- Open source components
- The operator, attack harness, and interceptor are published under AGPL-3.0-only for inspection and extension.
Good for
- Test whether a new tool or MCP integration gives an agent unintended authority
- Check that approval requirements and guardrails stop consequential actions
- Regression-test an agent after changes to its model, prompts, or tools
- Reproduce a past failure and verify that a fix closes the same path
- Evaluate how an agent behaves when its tools or dependencies fail
Intrusive.ai FAQ
Is the testing done against production systems?
No. The site says the surrounding world is replaced with observable, restorable simulations so exploration does not carry production blast radius.
Which license applies to the software?
The site states that the operator, attack harness, and interceptor are open source under AGPL-3.0-only.
What is recorded during a test?
The site says findings preserve the path from adversarial input to consequential effect, along with native observations and artifact digests.
Where this comes from
- Description: written from the project's website.