An AI safety layer designed for teens
Armadillo is a teen safety tool that layers on top of ChatGPT, Gemini, Claude, and Google Search's chatbot to make each of them better for teens. Armadillo is designed to close the loopholes in teen AI safeguards and to raise the bar for healthier teen chatbot use across all platforms.
Four risky things teens might ask
We ran Armadillo through KORA, a leading independent youth-safety benchmark, which tested 365 matched multi-turn conversations per arm for ages 13–17, graded by an external judge model as failing, adequate or exemplary. We tested ChatGPT 5.6 Luna with and without Armadillo's safety layer to try to simulate the experience of a non-logged in teen user.
When we added Armadillo, ChatGPT conversations graded as
failing fell from 24.4% to
3.8%,
a reduction of 84%.
The breakdown of grades for the chatbot responses across all 365 conversations, ages 13–17
On KORA's composite scale that is 72.1 for Luna, 81.2 for Sol, 88.8 for Astra, and 92.9 for Luna + Armadillo. Armadillo running on the cheapest GPT-5.6 tier beats both OpenAI's 5.6 Sol and flagship Astra model using OpenAI's own child-aware prompt.
Same corpus, same grading, each product in its own child mode.
Composite scores: 52.7 for SchoolAI, 86.4 for MagicSchool, 92.9 with Armadillo.
Right now, the best case scenario for a teen is that they close ChatGPT and go use MagicSchool instead. Armadillo makes the chatbot already open in their browser a safer option than the best purpose-built product on the board.
Share of conversations graded failing, ChatGPT 5.6 Luna with and without Armadillo. Lower is better.
Every key category improves meaningfully. Cognitive dependence (the risk of a teen outsourcing their own thinking) was ChatGPT's single worst category, failing every conversation in the set.
Benchmark: KORA kora-benchmark-tier3.4.1, 26 risk
categories, ages 13–17. Backing model
openai/gpt-5.6-luna, judge
gpt-5.2:medium:limited, simulated user
deepseek-v3.2. 365 matched conversations per arm;
both arms run the identical pipeline against the identical
scenario set, and pre-prompt injection was verified from the trace
files on every turn.
Where a composite score is quoted it uses KORA's own leaderboard formula, (Adequate% + 2 × Exemplary%) ÷ 2, applied identically to every arm. The difference between arms is significant at χ² = 63.58, p = 1.6 × 10−15; 20 of 26 individual risks improved and none worsened.
Third-party figures for SchoolAI, MagicSchool and GPT-5.6 Sol come from KORA's public age-filtered leaderboards, pulled separately from the Armadillo run. Results cover the 13–17 band only.
Each of these is a recorded, unaltered response from a chatbot. For each of four risky scenarios, you can compare ChatGPT, ChatGPT with a teen account, and ChatGPT + Armadillo to see the differences. Chatbot responses have been annotated to help break down the differences.
The teen types
1 of 12
The benchmark shows the prototype works against simulated teens. The next stage refines it alongside both experts and teens to validate both the science and the usability, before testing it in a large pilot study.
An advisory council of developmental psychologists, clinicians and youth-safety researchers sets the design principles, then we refine the instruction prompts against them and conduct expert-reviewed testing to ensure the product aligns with the principles.
A teen council co-designs the product features and results, before extensive beta testing of the refined product. By including teens from the beginning of the refinement process, we believe that we'll create a product that works better for teens.
Both tracks converge on a large-scale pilot test to help us measure the impact, before a round of final refinements. After this, the finished product is ready to go to market.
Leave your details and I'll reach out to connect with you.
If you'd rather just hear updates on Armadillo's progress, check the box and I'll add you to the update list instead.
Matt PuretzFounder, Armadillo
Enter your email and get early updates on Armadillo.