Stateful, Reversible Pseudonymization for Enterprise LLM Traffic.
DOST is a lightweight inline privacy proxy engineered to protect confidential enterprise data when employees and autonomous agents interact with cloud LLMs (ChatGPT, Claude, Cursor, Claude Code)—preserving full model reasoning and code generation integrity.
Destructive Redaction vs. Reversible Pseudonymization
Conventional DLP gateways replace sensitive strings with static [REDACTED] markers, destroying entity relationships and breaking generated code. DOST substitutes deterministic, type-consistent placeholders on outbound requests and restores original values on the inbound SSE stream.
Collapses distinct identifiers into identical [REDACTED] tokens. The upstream model loses referential context and returns non-functional code or ambiguous analysis.
where owner == [REDACTED]. Inbound LLM Response SELECT * FROM [REDACTED] WHERE id = [REDACTED];
// Referential mapping permanently lost
Maps sensitive entities (secrets, PII, Person IDs, codenames) to structured pseudonyms in an ephemeral session table, then translates tokens back in real time as SSE chunks arrive.
where owner == Person_ID_01. Inbound Stream to Client (Re-hydrated) SELECT * FROM kronos_prod.db WHERE id = 'TR-9948102';
// Restored inline with <20ms streaming overhead
The Four Pillars of DOST
Combines deterministic multi-pattern scanning (Aho-Corasick automata and Google RE2) with GLiNER entity extraction to identify candidate secrets, PII, and internal codenames while eliminating false positives.
Evaluates surrounding semantic context in milliseconds using quantized tiny ML models (GLiNER & GLiClass) to separate public technical terms from confidential enterprise data.
Replaces private attributes that must not leave the perimeter—such as Person IDs, credentials, and proprietary codenames—with structured, deterministic placeholders (e.g., Company_Alpha).
Reconstructs original values on inbound Server-Sent Events traffic via a sub-20ms sliding-window micro-buffer. Provider-agnostic and completely invisible to end users and coding agents.
Internal Appliance Pipeline
Detailed internal architecture of the 2-stage detection engine, ephemeral state vault, and inbound SSE stream re-hydration buffer.