Learna Tutor
An AI tutoring application for children aged 4 to 13 that lets children converse directly with an AI tutor. It incorporates a three-layer defense-in-depth safety architecture comprising an input filter, model system prompts and an output filter, with emphasis on privacy protection and academic integrity.
Timeline (last 90 days)
The children's AI tutoring app Learna Tutor is aimed at children aged about 4 to 13, allowing them to converse directly with an AI tutor.
Learna Tutor's safety design adopts a three-layer defense-in-depth architecture: input filters, model system prompts, and output filters.
Learna Tutor's input filter (checkTutorInput) contains eight categories of regex rules: privacy, crisis, bullying, medical/legal, out-of-scope, academic integrity, prompt injection, and unsafe web pages, and rejects unsupported tutor modes.
Learna Tutor's output filter (checkTutorOutput) has a dedicated privacy check; when the tutor asks the child for personal information, that response is blocked.
Learna Tutor's output filter handles negations correctly: "Don't share your email address" passes, because blocking safety reminders itself would be a false positive.
In the first round of red team testing, the prompt injection filter caught "ignore previous instructions" but missed "ignore your previous instructions".
The fixed prompt injection regex pattern is: ignore\s+(your\s+|all\s+|the\s+)?(previous\s+)?instructions, using optional capture groups to generalize the qualifiers.
After the fix, parallel coverage was added for disregard, forget, system prompt, jailbreak, developer message, and role-play escapes, the full suite was re-run and all passed, and six additional injection probes with new phrasings were added, all of which were blocked.
This round of red-team testing targeted only the filter layer and did not conduct end-to-end testing on the live model, because the test environment had no API key.
All Facts (9)
This round of red-team testing targeted only the filter layer and did not conduct end-to-end testing on the live model, because the test environment had no API key.
50%UnverifiedLearna Tutor's safety design adopts a three-layer defense-in-depth architecture: input filters, model system prompts, and output filters.
50%UnverifiedLearna Tutor's input filter (checkTutorInput) contains eight categories of regex rules: privacy, crisis, bullying, medical/legal, out-of-scope, academic integrity, prompt injection, and unsafe web pages, and rejects unsupported tutor modes.
50%UnverifiedLearna Tutor's output filter (checkTutorOutput) has a dedicated privacy check; when the tutor asks the child for personal information, that response is blocked.
50%UnverifiedLearna Tutor's output filter handles negations correctly: "Don't share your email address" passes, because blocking safety reminders itself would be a false positive.
50%UnverifiedThe fixed prompt injection regex pattern is: ignore\s+(your\s+|all\s+|the\s+)?(previous\s+)?instructions, using optional capture groups to generalize the qualifiers.
50%UnverifiedThe children's AI tutoring app Learna Tutor is aimed at children aged about 4 to 13, allowing them to converse directly with an AI tutor.
50%UnverifiedIn the first round of red team testing, the prompt injection filter caught "ignore previous instructions" but missed "ignore your previous instructions".
50%UnverifiedAfter the fix, parallel coverage was added for disregard, forget, system prompt, jailbreak, developer message, and role-play escapes, the full suite was re-run and all passed, and six additional injection probes with new phrasings were added, all of which were blocked.
50%