Introducing the PPQ Enclave Router. Every chat through PPQ now runs in an AWS Nitro enclave that PROVES: 1) PPQ cannot read the content of your queries 2) PPQ sent your query to the model and provider you chose. No model fraud. No more promises. Only provable guarantees. 🧡 image

Replies (26)

2/ Every AI router on the market can currently read the prompts passing through it. That's just how it works: your request has to be decrypted somewhere before it reaches the model, and that somewhere is the router's own servers. "We don't log your prompts" is only a policy.
5/ How it works: β€’ Your prompt is automatically encrypted to a key only the enclave holds β€’ The enclave decrypts it and sends it on to the provider β€’ The reply is encrypted back to you from the enclave β€’ PPQ's backend gets only billing and error metadata. No content!
6/ Amazon can't peek either. Nitro has no operator login, no SSH, and no way for any AWS employee to read enclave memory. AWS says this holds "including in fulfillment of a law enforcement request." NCC Group independently reviewed the design and found no gaps.
7/ Model fraud is ruled out the same way. The enclave's code binds each model family to its provider. A Claude request can go to Anthropic or the fallback of openrouter and nowhere else. Every response carries a signed routing receipt: the model you asked for, the provider it went to, the exact model id sent. Proven.
9/ We also shipped the PPQ Privacy Verifier: a small open-source proxy you run locally. Point any OpenAI- or Anthropic-compatible client at it. It fetches the enclave's attestation, checks it against the published build, and refuses to send on a mismatch.
10/ Add the rest of PPQ: no accounts, email optional, pay with Bitcoin, Lightning, Monero or stablecoins. The provider that can read your prompt doesn't know who you are. PPQ, which knows how you paid, can't read your prompt. Content and identity never sit in the same place.
11/ Honest limits: this covers chat today. Image, video and audio still hit our backend directly; bringing them in is next. Frontier providers like Anthropic still see your prompt, because that's where the model runs. For our E2EE models on tinfoil.sh, not even the provider does.
13/ PPQ exists to put the best AI in the hands of anyone, anywhere, safely and privately. No account, no subscription, no surveillance. The Enclave Router is the biggest step we've taken toward that goal yet. We will continue building in this direction forever.
14/ We believe the entire routing industry will eventually move towards enclaves. We are happy that we are among the first, but it is possible that security researchers may find some issues or points of improvement for our particular implementation. We welcome any and all feedback. Please reach out to us!
This move from promises to provable guarantees is the core of true trust. Receipts for model usage, like for zaps, turn claims into verifiable artifacts. It's the only way agents can truly coordinate.
The Nitro attestation angle is real: the verifier checks the cryptographic hash of the actual binary in memory, not a claim about what code is running, and the plaintext only exists inside the enclave where the parent instance can't capture it. That's what removes the operator from the trust equation.
Yes, that is true. And that's why we stick to premium providers such as Fireworks AI, Anthropic, Google, AWS, etc, for over 90% of our traffic. The big companies are very likely not going to defraud. It's all of the little shady providers running around that are more likely to do that. OpenRouter is kind of our generic backup for when the primary providers fail and also for the long tail of many lesser-known models, and for that, we rely on OpenRouter's own due diligence, which is decent but not perfect.
fade2's avatar
fade2 16 hours ago
Why is PPQ not releasing my referral credits until I use PPQ? This changed about 2 weeks ago. No notice. My refererred users are using PPQ and I'm not getting credited immediately. This is shady AF.
↑