AI-vendor promises
Self-attested policies with no external verification and no lineage. You're trusting a promise, not a proof.
GPT-Shield finds, blocks, or pseudonymizes sensitive data the moment your team sends it to ChatGPT, Claude, Copilot, Gemini — and 30+ more AI tools — then gives security a complete audit trail, without ever storing the secret.
Draft a follow-up to [EMAIL_1] about the overdue invoice. Bill account SSN [SSN_BLOCKED], and authenticate with key [API_KEY_BLOCKED].
Your people paste customer records, source code, and credentials into AI tools every day. Your DLP never sees it, your policies can't enforce it, and you have no record it happened. The tools you already own were never built for the AI boundary.
Self-attested policies with no external verification and no lineage. You're trusting a promise, not a proof.
Not AI-native, not real-time, not local, not boundary-aware. It never sees the prompt that matters.
Unenforceable, and it leaves no audit trail. One distracted paste and the secret is already gone.
The GPT-Shield Engine runs on your machine and makes the same decision whether the data comes from a browser, a desktop app, an internal API, or a file. Same input, same policy, same result — with no cloud round-trip to detect.
Sees through obfuscation and encoding tricks before anything is checked.
Validators, checksums & AI models spot 125+ types of sensitive data.
Scores severity and flags high-risk combinations — name + SSN = critical.
Block, redact, or pseudonymize — reversibly, without breaking the prompt.
Logs the crossing as proof — the shape and flow, never the value.
Five shields, one local engine. Together they cover the prompt going out, the file attached, the response coming back, and the trail it all leaves behind.
Stops a support rep pasting a customer record into ChatGPT.
Stops a PII-laden contract going into a RAG upload.
Flags prompt-injection, jailbreaks, and policy-bypass attempts before they leave.
policy-bypass attemptDetects reflected secrets and net-new sensitive data the model puts in its response.
reflected secret caught"entity_type": "national_id.us_ssn",
"destination": "openai",
"action": "block",
"raw_value_stored": false
Gives the CISO a full audit trail of what crossed, when, and where.
Core detection runs entirely on your machine — no cloud dependency. We never store raw sensitive values: only salted hashes and token references. Anything we do persist is encrypted with AES-GCM.
Read the full trust center →Detection needs no network. Your prompts never leave your machine to be scanned.
Salted hashes and token references only — the secret itself is never written down.
Anything persisted is AES-GCM encrypted, backed by the OS keychain.
Deterministic, auditable, and identical on every surface. Proof, not the secret.
Start with one team and a browser extension; expand to every surface under central policy. No rip-and-replace, no agents on critical paths you don't control.
Push to Chrome & Edge across the org. Covers ChatGPT, Claude, and Gemini.
Signed installers for Windows & Mac. Runs quietly in the tray for native AI apps.
Drop it in front of internal AI traffic and API gateways. OpenAI/Anthropic-compatible.
One policy engine, every surface. Expand coverage without re-training anyone.
DLP was never built for the AI boundary, and an AI vendor protecting your data from itself isn't protection. GPT-Shield is the only layer you control.
| Capability | GPT-Shield | Legacy DLP | AI-vendor controls |
|---|---|---|---|
| Sees the actual AI prompt | |||
| Protects before it reaches the model | |||
| Runs locally — your data stays put | |||
| Pseudonymizes (reversible), not just blocks | |||
| Audit trail of every AI crossing | |||
| Never stores the raw value | |||
| Browser, desktop, API & files | |||
| Works across every AI tool |
One platform, three jobs done — control for the CISO, rollout for security engineering, evidence for compliance.
A complete, payload-free audit trail of every AI crossing — evidence for the board and auditors, not a vendor's promise.
One policy engine across browser, desktop, and gateway — boundary-aware rules, tunable false positives, no agents on paths you don't control.
Local-first processing and proof a sensitive value never left the building — ready for regulators, DPAs, and data-residency requirements.
No. Core detection runs locally on each machine — prompts are scanned in place and never sent to us to be checked. Optional cloud features only ever receive de-identified signal — classifications and counts — never raw values.
Detection on a typical prompt is sub-millisecond and happens on-device, so there's no cloud round-trip in the critical path. Users won't feel it.
Policies are boundary-aware and tunable — a value allowed for an internal model can be pseudonymized for a public one. Reversible pseudonymization means even a cautious redaction never destroys the original.
Yes. Pseudonymized values map to stable tokens and can be restored for trusted destinations. Blocked values (like credentials) are removed and stay removed.
Yes. The gateway is self-hostable and OpenAI/Anthropic-compatible, and the engine runs entirely on your infrastructure. There is no required cloud dependency for protection.
ChatGPT, Claude, and Gemini in the browser today, native desktop AI apps via the agent, plus any internal AI or API traffic through the gateway, and file uploads.
We built GPT-Shield because the fastest-moving risk in every company is also the hardest to see: a teammate pasting something sensitive into an AI tool to get their job done. You can't train your way out of it, and you shouldn't have to choose between moving fast with AI and protecting your data.
So we built the layer in between — one that runs on your machine, blocks or pseudonymizes what matters, and proves what crossed the line without ever keeping the secret. That principle isn't a feature; it's the whole point.
See it on your own data in a 20-minute demo — or grab an early-access invite.
No raw data leaves your machine. Ever.