Is Janitor AI Safe? Why "Who Runs the Model" Is the Real Question

AI companion apps split into two structural categories based on who runs the underlying model.

Bring-your-own-model platforms provide the chat interface (character cards, prompts, chat UI) but ask you to plug in access to an AI model separately — either your own paid API key (OpenAI, DeepSeek, Anthropic) or a third-party reverse proxy that shares one. Janitor AI is in this camp.

Full-stack platforms run their own model access, their own billing, and their own moderation. You sign up, you chat. There's nothing to configure between you and the AI. Sloane is in this camp.

Both shapes are legitimate. Which one is safer depends on which safety questions matter to you — and the questions are genuinely different, not just "one is better."

On a BYO-model platform, the safety questions are mostly about the pieces you connect, not the platform itself.

The reverse proxy question. A reverse proxy is a third-party service that stands between the chat platform and a paid AI model API. Users use them because sharing one paid subscription via a proxy is cheaper per-message than each user paying full price. The catch: the proxy operator sees every message that passes through, can log conversations, and can extract API keys that transit the proxy.

Community-shared "free" proxies are the specific risk. If someone posts a proxy URL in a Discord and hundreds of users hit it, the operator is one screen-scraper away from having every message from every user. There's no accountability, no company to sue. The proxy operator is often literally anonymous.

The API key question. If you use Janitor with your own API key instead of a shared proxy, the risk shifts. Leaked OpenAI keys with no rate limit set can generate hundreds of dollars of charges in an afternoon. Mitigation is straightforward — set hard monthly spending caps at the provider (OpenAI, Anthropic, and DeepSeek all support per-key monthly limits), and never paste keys into untrusted proxy interfaces.

The fake-site question. Because Janitor is popular and its architecture is straightforward to clone, fake sites impersonating Janitor branding show up regularly. The legitimate Janitor is `janitorai.com`. Anything else (janitorai.net, janitor-ai.com, subdomain variations) should be treated as suspect until you've verified it independently.

These are real risks, but they're the *specific* risks that come with the BYO-model architecture. For a user who's comfortable managing API keys and specifically wants the "no company can shut this off" property that comes from user-owned model access, these are workable tradeoffs.

On a full-stack platform, the "who runs the model" and "who handles billing" and "who moderates" questions all collapse into one: it's the vendor. The safety questions become the standard SaaS safety questions — named operating entity, standard payment processor, self-serve cancel, documented privacy commitments.

Sloane runs its own model access, its own billing (Stripe, standard SaaS), and its own moderation. There's no proxy to evaluate, no API key to leak, no third-party service standing between you and the persona. The safety story is: one company's privacy and platform commitments, transparently documented, one place to escalate if something goes wrong.

The tradeoff of the full-stack shape is that you're trusting a single vendor rather than being able to swap models yourself. That's the right tradeoff for most users (fewer moving pieces, no configuration risk) and the wrong tradeoff for the specific "I want maximum control" audience Janitor serves. Both shapes are legitimate — the question is which set of safety concerns you'd rather manage.

Kaya is a good default if you want to see the full-stack version of the category. Free tier is real — 50 messages/day with any persona in the roster, no card at signup, real conversation with real memory, cancel via the Stripe Billing Portal if you upgrade and later change your mind.

If you've decided Janitor is right for your needs — the BYO-model shape fits how you like to work — here's the specific safety hygiene worth running.

Verify you're on janitorai.com. Not janitorai.net, janitor-ai.com, or any subdomain variation. Fake sites are the fastest way to lose an API key.

Use API keys with hard spending caps. OpenAI, Anthropic, and DeepSeek all support per-key monthly limits. Set them low relative to what a leak could cost — $20-50/month cap is enough for personal chat and hard-blocks any runaway leak.

Never paste your personal API key into an untrusted proxy. If you can't read the proxy's code or verify its operator, don't send your key through it. Once a key transits a compromised proxy, treat it as leaked and rotate.

Prefer official APIs over community proxies where possible. Using OpenAI's or DeepSeek's official API directly costs slightly more per message than a shared proxy, but you get the accountability, rate limits, and audit trail that come with using the real provider.

Rotate keys immediately if you suspect a leak. All three main providers have one-click key rotation in the dashboard. Rotating is free and takes seconds.

Run this hygiene and Janitor is safe to use in the specific way it was designed to be used.

One caveat worth naming that applies regardless of whether the platform is BYO-model or full-stack: conversations with any AI companion product become model inputs somewhere in the chain, and privacy claims should be read specifically rather than trusted broadly.

On Janitor, your conversations pass through the model provider (OpenAI, DeepSeek, whoever) and potentially through a proxy operator too — each of those has its own privacy policy that governs what happens to your data. On Sloane, conversations pass through Sloane's own infrastructure and the underlying model provider Sloane uses — one privacy policy chain to read, documented here.

Behavioral mitigation is the same regardless of shape: don't share personally identifying information (real name, address, workplace, financial details, specific medical details) in AI conversations you wouldn't want stored somewhere. That's baseline hygiene for any AI product, not paranoia.