Home›Read›Tools desk›GLASTOR releases pii-veil claiming sub-millisecond reversible PII redaction in Rust
Tools·Nov 19, 2026

GLASTOR releases pii-veil claiming sub-millisecond reversible PII redaction in Rust

GLASTOR's open-source Rust library pii-veil claims 165 µs PII redaction and stateful SSE stream restoration for LLM prompts, positioning itself against Python-based Presidio and cloud DLP APIs. For…

GLASTOR's open-source Rust library pii-veil claims 165 µs PII redaction and stateful SSE stream restoration for LLM prompts, positioning itself against Python-based Presidio and cloud DLP APIs.

For infrastructure teams building inline LLM middleware or Model Context Protocol (MCP) agents in Rust or WASM, pii-veil offers a lightweight, zero-egress alternative to heavy Python NLP pipelines and cloud DLP services. Skip it if your application stack relies on deep contextual entity extraction that requires spaCy or custom spaCy models. GLASTOR's open-source engine trades broad statistical NLP models for deterministic speed and stateful SSE chunk restoration.

Methodology

This v0 review draws on the published technical post and benchmark claims from GLASTOR (observed November 2026) at https://dev.to/andres_003352671e0820b880/how-we-built-a-165-us-reversible-pii-anonymizer-for-llms-mcp-in-rust-212x-faster-than-presidio-25a0. Independent benchmarks remain pending. Update cadence: re-tested when claims diverge from observed behavior.

The scope of this evaluation covers GLASTOR's published criterion benchmark claims, technical architecture descriptions for StreamingRestorer, live WASM playground behavior at https://pii-veil.pages.dev, and source implementation details in the glastor-core/pii-veil repository. This review does not cover independent reproduction of cargo bench -p pii-veil-core --bench bench_throughput, long-term memory safety under high-concurrency stream pressure, or edge-case entity identification accuracy against complex unstructured text payloads.

What it does

GLASTOR built pii-veil to address data privacy for LLM prompts, RAG pipelines, and Model Context Protocol (MCP) agents without sending raw text to third-party endpoints. Instead of relying on Python processes running Microsoft Presidio with spaCy (~35 ms latency, ~450 MB RAM footprint) or Cloud DLP APIs like AWS Comprehend PII (80-250 ms latency, $100+ per million requests), pii-veil executes inline in Rust.

Pattern matching in Rust

The core engine redacts sensitive patterns including email addresses, credit card numbers, IBANs, national tax IDs, PostgreSQL connection strings, and leaked API keys. In GLASTOR's reported Criterion benchmarks, scanning a 1 KB prompt containing 7 entities takes 165.68 µs (0.165 ms), while single-entity scans register at 4.01 µs. Throughput for a 1 MB buffer is reported at 20.58 ms (~46.3 MiB/s).

Stateful SSE chunk restoration

The key architectural addition in pii-veil is StreamingRestorer, a state-machine buffer designed for Server-Sent Events (SSE). When LLMs stream responses token by token, placeholder tokens like [EMAIL_1] can be split across SSE payload boundaries (such as [EMA in chunk 1 and IL_1] in chunk 2). StreamingRestorer tracks partial placeholder prefixes at chunk boundaries to reconstruct original entities in a reported 69.65 µs per SSE chunk without requiring full-response buffering.

Client-side WebAssembly execution

GLASTOR packages pii-veil for WebAssembly targets, allowing browser-based client-side redaction where zero data leaves the client device, accessible via their live playground at https://pii-veil.pages.dev.

What's interesting and what's missing

The architectural highlight of pii-veil is tackling SSE stream restoration as a state machine problem rather than a post-processing string replace. Traditional redaction solutions either mandate full-response buffering before rendering tokens to end users, destroying the perception of streaming speed, or fail entirely when boundary tokens split placeholders. Achieving 69.65 µs per chunk restoration latency keeps inline proxy overhead well under a single animation frame.

However, the comparison against Microsoft Presidio requires context. Presidio uses spaCy named entity recognition (NER) models alongside regex patterns, allowing it to detect context-dependent entities like names and addresses that lack deterministic structures. GLASTOR's 212x speedup claim against Presidio (~165.68 µs vs ~35 ms) compares a high-throughput compiled Rust engine against a Python statistical NLP pipeline. If pii-veil relies primarily on regex and deterministic token matching, its raw scanning throughput will naturally outpace spaCy, but it may miss ambiguous entities that require statistical context. GLASTOR has not yet published comparative recall and precision benchmarks against standard PII evaluation datasets.

Pricing

pii-veil is released as open-source software under GLASTOR core repositories on GitHub as of November 2026. The core Rust library and WASM binaries are free to inspect, self-host, and embed. Cloud DLP pricing comparisons cited by GLASTOR reference AWS Comprehend at $100+ per million requests.

Verdict

pii-veil is an ideal fit for Rust-native AI microservices and MCP proxy nodes requiring sub-millisecond redaction without adding cloud DLP network costs or Python sidecar memory overhead. For teams handling structured secrets, credit cards, emails, and database URIs in high-throughput streaming interfaces, GLASTOR's stateful SSE buffer solves a genuine streaming UI problem. Teams requiring complex natural language entity detection (such as arbitrary personal names in conversational prose) should verify whether pii-veil's rule set covers their domain before replacing full NLP stacks like Microsoft Presidio.

What we'd test next

We plan to run independent cargo bench tests across multi-threaded loads to verify the reported 46.3 MiB/s throughput. We also intend to benchmark StreamingRestorer latency under synthetic SSE stream fragmentation patterns, deliberately splitting placeholders across three or more single-character chunks. Finally, we will measure entity detection recall against the CoNLL-2003 and Microsoft Presidio standard test suites to quantify the precision trade-off of Rust pattern matching against spaCy NLP pipelines.

The investor read

GLASTOR's launch of pii-veil signals a broader migration of AI infrastructure primitives from Python sidecars and third-party SaaS DLP APIs to compiled, zero-egress Rust and WebAssembly modules embedded directly into API gateways and MCP agent routers. As enterprises deploy AI agents handling sensitive customer data, latency overhead and cloud DLP costs ($100+ per million calls on AWS Comprehend) become immediate friction points. Tools like pii-veil demonstrate how inline edge processing can drastically reduce compliance overhead. As an open-source utility, pii-veil functions primarily as a technology demonstrator and adoption wedge for GLASTOR's platform. To become venture-scale infrastructure, GLASTOR will need to monetize enterprise governance control planes, audit logging, and automated compliance policy distribution across distributed agent networks.

Pull quote: “The key architectural addition in pii-veil is StreamingRestorer, a state-machine buffer designed for Server-Sent Events (SSE).”

Sources · how we verified
  1. How We Built a 165 µs Reversible PII Anonymizer for LLMs & MCP in Rust (212x Faster Than Presidio) ↗

Every claim ties to a primary source. See our methodology.

Reported by the Riley desk on Founderr Pulse’s Tools beat. Every factual claim is tied to a primary source and linked; anything that can’t be stood up doesn’t run. Founderr (RIKHATH LLC) is the accountable publisher and corrects in place. How we work · About · File a correction.
R
Riley

The Riley desk covers tools — what founders are building with, switching to, and abandoning. Every claim is sourced and linked. Operated by Founderr (RIKHATH LLC) See the desk →

Founderr Pulse — free & independent. The desk for people who build & back.