ScaryBench
ScaryBench is an educational service that provides an introduction to the practice of AI alignment research. It allows you to work with a misaligned AI agent and use the software tools employed in alignment evaluations.
Here you can:
- Launch an evaluation of a malicious AI agent, which runs inside of a sandboxed environment.
- Watch as that agent researches you on the internet and creates a phishing email and website designed to trick you into giving up private information.
- Review the results of the evaluation and the full log of the agent's actions in Inspect, the software that alignment researchers use to test AI models.
If you just want to see what Inspect looks like, you can view the results of some previous ScaryBench runs here.
To run an eval that simulates a phishing attack against you, sign in with LinkedIn below. Already signed up? Go here to get back to your run. You can always change the visibility of your results from public to private after the run is complete.
FAQ
What AI model are you using in the evaluation? We are currently using DeepSeek-V4.1-Flash.
What prompts are you using? You can see the full set of prompts by reviewing the run history in Inspect.
How is the agent contained? The agent can't run code — it only gets a fixed set of browser tools. It can navigate to public sites to research you (internal networks are blocked), but it can only type and submit forms on search engines and LinkedIn — nowhere else. The phishing email it writes is displayed here, never actually sent. Since it reads the live web and publishes a real, public page, treat anything it creates as public.
Why is it called "ScaryBench?" The "bench" is short for "benchmark." This is a standard naming convention in AI evals. For example, LiveCodeBench is a set of tests that benchmark a model's capability to write code. Our tests benchmark, in part, how scary an agent is.
What if I have more questions? Email hello@scarybench.org.