Thursday, November 19, 2026

10:30 AM - 11:30 AM

Swan Room

---

Presentation Outline

You want to use AI — and you have been told no. The no was right: a language model reads everything it can reach, a cloud GPU holds your plaintext in memory you do not control, and even a search index can betray that a classified document exists without returning a word of it. But every no is anchored to a disclosure path, and every path now has an engineering answer you can test. The blocker is no longer "unsafe". It is "unfamiliar". The session takes the six no's apart, one wall at a time. "The model will read everything" — not if access rules are enforced at the data itself, binding a model to clearance and caveats exactly as a person is bound. "Your plaintext sits on silicon you don't own" — processors and GPUs now cryptographically prove what they are running — hardware attestation — AMD SEV-SNP and NVIDIA's confidential-computing modes among them; the proof is checked locally, air-gap ready. "One system can't serve every classification" — so route each request to an engine trusted for its level, and refuse when none is. "Even the search leaks" — then build search that cannot disclose what it refused to return. "The keys have to go somewhere" — they never leave your boundary. "It won't survive audit" — nothing releases without a record, and no record can be quietly rewritten.

---

Speaker