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 things teens ask
We ran Armadillo through KORA, an independent child-safety benchmark: 365 matched multi-turn conversations per arm, ages 13–17, graded by an external judge model as failing, adequate or exemplary. The only variable between arms is whether Armadillo is running.
Conversations graded failing fell from 24.4% to 3.8% — 89 failures down to 14, an 84% reduction.
How the judge graded every conversation, ages 13–17, 365 conversations per arm.
On KORA's composite scale that is 72.1 for Luna, 81.2 for Sol and 92.9 with Armadillo. Armadillo running on the cheapest GPT-5.6 tier beats OpenAI's flagship 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 for a teen is that they close ChatGPT and go use MagicSchool instead. They mostly don't. 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 category here drops by more than 60%. 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 real exchange, screenshot as it happened. For every scenario you can look at plain ChatGPT, ChatGPT with a teen account, and ChatGPT with Armadillo running on top of it. The teen account changes the tone far more often than it changes the behavior.
The teen types
1 of 12
The benchmark shows the prototype works against simulated teens. The next stage tests it against real ones, on two tracks that feed each other.
An advisory council of developmental psychologists, clinicians and youth-safety researchers sets the design principles, then we refine the instruction prompts against them and re-run the benchmark. This is the track that decides whether Armadillo's behavior is defensible, not just measurable.
A teen council co-designs the interface and beta tests the full prototype. Teens are the people who can uninstall it, so this track decides whether any of the above survives contact with the person it was built for.
Both tracks converge on a large pilot test, final refinements, and a shipped product.
Leave your details and I'll email you to set up a time. If you'd rather just hear how this progresses, check the box and I'll add you to the update list instead — no follow-up from me.
Matt PuretzFounder, Armadillo
Leave your email and I'll reach out to set up a meeting.