AI-agent egress redaction

Your agents read secrets. Stop them leaving.

ShroudCloud sits on the egress path of your AI agents — every tool result they read, every prompt they send — and masks credentials and PII before the model or the network sees them. Deterministic placeholders keep the work going; the real values never leave your boundary. Nothing retained, ever.

CISSP® certifiedRunning on our own agent stackResearch at cctbp.comOpen lab: FlawedToken

Where the guard sits

Agents leak on egress, not on access. A value is read into context and then walks out — in the next prompt, in the next tool call, in the reply to chat. ShroudCloud covers all three exits.

1 · Inbound: tool results

A hook in the agent masks each tool result — shell output, file reads, web pages, API responses — before it enters model context. This is the path where credentials arrive uninvited.

2 · Outbound: prompts to the model

The guard runs as an OpenAI-compatible endpoint beside the agent. Point your client's base URL at it and every prompt is masked and sealed before it reaches the provider.

3 · Outbound: agent replies

The message your agent posts to chat is filtered on the way out, and placeholders are never restored on this path — restoring here would put the real value in front of a third party.

What it catches

Detection runs on your side of the boundary. Each of these was live-probed against the engine before it went on this page.

EntityWhy it matters for an agent
API keysAWS access keys, GitHub tokens, OpenRouter / Anthropic / OpenAI keys, HuggingFace tokens, Telegram bot tokens — the credentials that turn one bad tool result into a rotatathon
Identity artifactsJWTs, bearer tokens, OAuth codes, session cookies — the things an identity researcher knows leak
PIINames, emails, phone numbers, addresses, dates of birth, US SSNs, credit cards, IP addresses
Generic secretsHigh-entropy strings in key/value context, private key blocks, PEM bodies

Profiles decide the mix: a secrets-only profile for agents that only need credential masking, a wider profile when PII is in scope, a strict profile that also takes network identifiers.

Three doors, one engine

Same detection engine, three different things you might be guarding. Pick the one that matches what you have in front of you.

Agents

A live agent session

Tool results, prompts, and replies masked as they move — so an agent that reads a credential cannot pass it on. This is the page you want if your problem is "the agent leaked it".

See the agent egress guard
Documents

A file you have to hand over

Upload a PDF or Word file, review every proposed redaction, then export a verifiably clean document with a complete audit trail. Built for legal, HR, insurance, and records work.

Explore document redaction
Engine

You're building it yourself

Run the engine where you want it: POST text, get deterministic placeholders back, keep the mapping on your side. Self-hosted, stateless, nothing retained.

Explore the engine API

Why ShroudCloud

Built for the agent path.

Most tools redact a document. ShroudCloud redacts the stream an agent lives in — tool results, prompts, replies — because that is where the leak actually happens.

Nothing to leak.

Nothing retained, by design. Text is processed, verifiably masked, and destroyed. Masked values are recorded as hashes, the reveal key stays on your side, and nothing is ever used for training.

A person, not a logo.

Built by a CISSP-certified identity security researcher who publishes the research behind it — the same research that runs the open-source FlawedToken lab.

Want this in front of your agents?

Tell us what your agents touch — models, tools, data sources — and we'll point you at the right shape of the guard. Early access users set the roadmap.

No pricing announced yet. Early access users lock in founding terms.