LAST UPDATED 2026-08-17
Trust & Security
An unfiled invention is a trade secret. This page states exactly how it is protected here — including the parts that are necessarily not encrypted, and who can see what. If a sentence on this page ever overstates what the software does, that is a bug; report it like one.
Your content is encrypted above the database
Everything that reveals your invention — idea descriptions, titles, intake details, drafting inputs and outputs, uploaded materials, and every generated document — is encrypted with AES-256-GCM before it reaches storage. This is in addition to the storage layer's own disk encryption and TLS in transit, and it is the stronger promise: someone holding a copy of the database, or a leaked storage credential, holds ciphertext.
Each account has its own random 256-bit data key. Data keys are wrapped by versioned master keys that live only in the runtime environment — the database never sees them, so the database alone can never decrypt its own contents. Every ciphertext is bound to the exact record it belongs to, so a value copied between rows or accounts refuses to decrypt.
What stays readable, and why
Operational fields remain plaintext because the product cannot run without querying them: statuses, timestamps, run costs and token counts, model names, and your certified entity size (micro, small, or large). None of these reveal what your invention is — they reveal that a draft exists, when it ran, and what it cost.
If the service refuses an action citing missing encryption keys, that is deliberate: the software fails loudly rather than ever storing your content unencrypted.
What AI providers see during drafting
Drafting is performed by large language models. While a draft is being generated, your idea and the text derived from it are necessarily processed in plaintext by the model provider serving that step — Google (Gemini) by default, or OpenAI or Anthropic where your account selects those models — over encrypted connections, under paid API terms whose providers state that API content is not used to train their models.
This is the honest boundary of the encryption story: no drafting tool can have a model write your specification without showing it the invention. What we control is everything before and after that moment — and we encrypt all of it.
Who can access your content
You, signed into your account; anyone you deliberately share a package with through a share link (which expires and can be revoked); and the service itself while performing the work you asked for. Download links are short-lived, bound to a single document, and served decrypted only after an ownership or share check.
There is no admin view of your invention text. Operational support sees the plaintext fields described above — statuses and costs — not content.
Deletion is cryptographic
Deleting a patent removes its records and documents. Deleting your account also destroys your wrapped data key — after which every remaining ciphertext belonging to you, including any backup copy anywhere, is permanently unreadable. This is crypto-shredding: deletion that does not depend on finding every copy.
Why this matters for patent rights
Until you file, confidentiality is not just privacy — it is what keeps your invention a trade secret and avoids a public disclosure that could start legal clocks or forfeit foreign rights. Keeping your materials encrypted, unshared, and undisclosed until the day you file is part of the product working as intended. Filing decisions and disclosure questions are legal judgments; for those, talk to a registered patent attorney or agent.
Questions and reports
Security questions or suspected vulnerabilities: security@avelith.ai. If you believe any statement on this page overstates what the product does, include the sentence — that report is treated as a defect.