Employees use websites outside the API gateway
Extend request policy to supported ChatGPT and Claude web editors without intercepting TLS or exposing provider credentials.
ChatGPT and Claude browser security
Protect prompts on ChatGPT and Claude before employees send them.
Entrovik Web AI Guard applies the active enterprise request-policy pipeline to visible prompts and locally extracted text attachments, with optional browser-response inspection disabled by default.
Why teams act
AI adoption creates a new control surface. Put enforcement in the path.
Entrovik gives security, platform, compliance, and application teams a shared technical boundary without coupling enterprise policy to one provider SDK.
Sensitive data leaves in prompts and attachments
Evaluate visible prompt text plus supported text, source-code, JSON, CSV, XML, YAML, log, and configuration files before submission.
Public AI usage lacks enterprise context
Attach an authenticated Entrovik identity and application label to each browser decision and produce content-minimizing audit evidence.
Enforcement model
Central policy with application-level context.
Every decision can use normalized content plus tenant, identity, team, application, provider, model, stage, and configured policy metadata.
01Pre-submit enforcement
An isolated content script pauses supported send actions while Entrovik evaluates visible editor and extracted attachment text.
02Local bounded extraction
The browser hashes and extracts supported text files locally; PDF, Office, image, binary, unreadable, and oversized files fail closed.
03Optional response quarantine
Default-off response inspection can withhold newly generated content until the tenant response pipeline allows or governs it.
04Credential isolation
The Entrovik credential stays in trusted extension storage and policy calls run from the Manifest V3 service worker, outside the website page.
In the request path
One decision lifecycle. Every interaction.
Policies remain modular packages. The core resolves, executes, contains, explains, measures, and audits them.
- 01
A managed content script pauses Enter, click, or form submission and captures selected files
- 02
The browser extracts only bounded supported text and sends text plus file metadata through the isolated service worker
- 03
Entrovik runs the tenant's active request pipeline without calling an AI model
- 04
A denial keeps the prompt or attachment unsent; prompt redaction is written back before submission
- 05
Optional response inspection, when enabled, quarantines new assistant content until response policy decides
Technical FAQ
Questions teams ask before deployment.
Does the extension read ChatGPT or Claude passwords?
No. It does not receive website passwords, session cookies, or provider credentials. It authenticates separately to Entrovik from its service worker.
Does it inspect AI responses?
It can, but browser response inspection is disabled by default. When enabled, new responses remain quarantined until the active response pipeline allows, blocks, or governs their display.
Does incoming PII or PHI automatically get blocked?
No. Incoming sensitive data is not inherently a violation. Response policy is useful when connected data could reach an unauthorized viewer or a deployment has explicit display controls.
What happens when Entrovik is unavailable?
The recommended and default behavior is fail closed, so submission remains blocked. Administrators may explicitly choose fail open for lower-risk deployments.
Which attachments can it inspect?
It locally extracts bounded text, code, Markdown, CSV, JSON, XML, YAML, logs, and related formats. PDF, Office, image, binary, unreadable, and oversized files remain blocked until trusted extraction or OCR is available.
Can users disable it?
An ordinary extension can be disabled. Enterprise enforcement requires browser-management policy, restricted unmanaged access, distinct device or user identities, and selector-health monitoring.
Put it in front of a real workflow
Turn your AI policy into an enforced decision.
Bring one provider path, one policy requirement, and the architecture your reviewers need to trust.