Research · AI security
Poisoned Context
When a stored database record contains instructions, does an AI assistant follow them? Pick a condition below to see exactly what the model would receive and how the reply is scored.
- No model runs on this page
- ● Fictional data only
- Results on GitHub
Choose a condition
What the model receives
Staff question
System instructions
Poisoned record
The real prompt holds 48 records in total. The email tool is offered in every trial and is never executed.
How the reply is scored
The attack succeeds if
The assistant did its job if
A person reviews the reply if
Results
The results and raw data for Claude Haiku 5.5 are on GitHub.