Security

The architecture, written like a researcher writes it.

Every redaction product makes a trust claim. This page is where we show our work: exactly what is logged, what is not, how data moves through the system, and what happens at every step. It will be kept current as the system evolves — this is a living document, not a marketing page.

Retains nothing: the standing commitment

Files and text are processed in memory and destroyed. No PII is stored server-side, and nothing is ever used for training. Only non-PII metadata (timestamps, entity counts, durations) is logged for the audit trail. The processing pipeline is designed so that a breach of the platform yields no customer content — there is nothing to exfiltrate.

What we log (and what we refuse to log)

  • Logged: request timestamps, entity counts, processing duration, error codes. Counts and metadata only — never content.
  • Never logged: document contents, extracted text, entity values, placeholders, or any mapping between them.
  • Retention: files and text are processed in memory and destroyed — nothing is retained in plaintext; a value you choose to make recoverable is held only as ciphertext under your key.

How data moves

Text and files are processed in memory by the detection engine in an isolated container. The verified output is returned to the caller; nothing is stored between jobs. The engine speaks JSON in and out — no persistent state between jobs.

The agent path

When the guard runs in front of an agent, the flow has one extra property worth stating: the engine holds no key and no plaintext.Text is sent, masked spans come back, and the masked values are recorded as hashes. Original values are sealed locally against the reveal vault's public key, so restoration happens on your side of the boundary — never on a path bound for a third party. The guard covers the paths an agent actually uses (tool results it reads, prompts it sends, replies it posts), and its limits are published rather than implied: no TLS interception for clients you cannot configure, no pattern detection for short or opaque secrets, and whole-request processing rather than streaming.

Fail-closed verification

The export pipeline re-runs detection on every output file. If any detected entity survives, the export fails closed — no partially redacted file ever leaves the system. Copy-paste, metadata, and programmatic text extraction are all tested as part of the release gate.

The threat model, honestly

The primary risk to a redaction product is a leak: a single missed entity in an exported file. That is why correctness is the product, not a feature. The secondary risk is platform compromise; retaining nothing is the mitigation, because there is nothing to take. We apply WAF rules, rate limiting, and least-privilege infrastructure as baseline posture.

Built by a researcher

The detection engine is informed by the same research published at cctbp.com and the open-source FlawedToken lab — a practitioner's understanding of where identity artifacts actually hide in real systems.

Where else the engine sits

Zero Trust isn't five walls, it's one posture. The same engine also goes where the data already flows in your stack.

PillarWhere ShroudCloud redacts
IdentityAuth and SSO audit logs, IAM telemetry, session records — JWTs, OAuth codes, bearer tokens
DevicesEndpoint telemetry, MDM inventory data, device logs before they hit the lake
NetworksDNS, NetFlow, email gateway logs — metadata and payloads stripped of identifiers
ApplicationsAPI payloads, support tickets, RAG pipelines, and LLM prompts before the model sees them
DataDLP pipelines, documents, databases — masking, verified-clean export