Show HN: Intrusive.ai – open-source AI attack simulation
Through my own experience with evaluating agents, I found indirect prompt injection testing to be especially difficult due to the architectural nature of it. To overcome that challenge, I created Interceptor Sandbox - a sandbox that runs agents in a simulated environment where all of the services an agent might rely on (HTTP, MCP, Postgres/pgvector to name a few) can be "intercepted" via an injection API to inject malicious payloads.
Interceptor Sandbox: https://github.com/intrusiveai/interceptor_sandbox
Additionally, I found that existing open source AI testing and attack simulation solutions didn't provide the breadth of capabilities I needed so I created Attack Harness for that purpose. But it also had to be sandboxed because any sufficiently capable attack harness can't be trusted to play nice once it's been given nefarious goals and so that led to Operator Sandbox. I haven't seen other people talking about the sandboxing method that I've employed in Operator so it seems to be a somewhat novel approach that might shake up the way people are thinking about sandboxing AI agents.
Operator Sandbox: https://github.com/intrusiveai/operator_sandbox
The full suite is documented here: https://docs.intrusive.ai
The platform is still very much in its alpha stage so your mileage may vary but I would sure appreciate any feedback that comes my way. Feel free to use GitHub Issues on the respective repos to provide feedback.
There are no comments yet.