Protect your company’s information when using AI.
Cortex Guard helps your organization detect prompt-injection risks in files and skills. With browser protection enabled, it can also check files and messages sent through supported AI websites and apply your organization’s rules for sensitive information.
What is prompt injection?
Files, documents, skills, and pasted text can contain instructions intended to manipulate an AI assistant. These attacks are known as prompt injection.
A malicious or compromised document might try to get an assistant to:
- Ignore its original instructions.
- Reveal confidential information.
- Misuse connected tools or attempt unsafe actions.
- Send information to an unauthorized destination.
These instructions can be mixed into ordinary content or hidden from view. A document that looks trustworthy can still contain text aimed at the AI rather than the person reading it.
What Cortex Guard checks in TextCortex
Depending on your organization’s settings, Cortex Guard checks new or changed content in these categories:
- Uploaded files: external files added to TextCortex.
- Skills: skill instructions and supported package content.
- Synchronized files: content brought in through enabled file-sync connections.
These checks look for prompt injection before enrolled content becomes available for AI use. They apply to content owned by users who hold an active, non-trial seat in the organization covered by Cortex Guard. They aren’t a promise that every existing file has been scanned.
Understanding content scan results
- Pending: the content hasn’t been cleared. A scan may be waiting, unable to finish, or held because a subscription or seat requirement needs attention.
- Scanning: the security check is running.
- Clean: no prompt-injection risk was detected by the scan.
- Quarantined: a possible prompt injection was detected. The content stays unavailable unless an administrator explicitly releases it.
- Released: an administrator has approved the scanned revision for use.
Enrolled content stays unavailable while it’s pending, scanning, or quarantined. A pending result doesn’t necessarily mean the content is malicious.
Clearance applies to the checked version of the content. Changing a file or skill invalidates its previous clearance.
Protection on supported AI websites
Company information can also leave through files and prompts sent to third-party AI tools. Cortex Guard’s browser protection checks supported submission flows on ChatGPT, Claude, and Gemini through the TextCortex extension for Chrome and Edge.
Access requirements: Browser protection needs a compatible extension, a signed-in account with a covered seat, and an enabled browser policy. Your administrator controls which supported sites and checks are active.
Depending on that policy, browser checks can look for:
- Prompt-injection instructions.
- Hidden instructions in PDFs, including text that readers can’t see.
- Supported personal-data patterns, such as email addresses, phone numbers, IBANs, card numbers, and certain national or tax identifiers.
- Recognizable secrets, such as API keys, private keys, and connection strings.
- Confidential-topic keywords configured by your administrator.
The personal-data, secret, and topic checks described here belong to browser protection. They shouldn’t be assumed to run on every file stored in TextCortex.
What happens to browser findings?
Browser findings don’t put all result in quarantine. Your organization sets a mode for each supported site:
- Off: Cortex Guard doesn’t check the site.
- Monitor: findings are reported without blocking the submission.
- Enforce: configured detector rules can allow, warn, redact, or block content.
When redaction is available, you can replace detected values before sending. A warning may allow you to continue with a reason if your organization permits it. For blocked content, remove the item, edit the message, or request administrator review when that option is available.
If a browser check can’t finish, your organization’s failure policy determines whether sending is stopped or allowed. An incomplete check isn’t a clean result.
What to do if content is blocked
- Read the message to understand whether the content was flagged or the check couldn’t finish.
- Remove suspicious instructions or unnecessary sensitive information when you can do so safely.
- If you think the result is incorrect, contact your administrator. For eligible browser findings, use the review-request option and explain why the content is needed.
Administrators manage stored-content scans in Tenant Admin settings, Cortex Guard, TextCortex content. They can retry eligible incomplete scans or explicitly release a quarantined revision. Browser review requests are handled separately in the Review requests tab.
Coverage and limitations
Cortex Guard checks supported content and the text it can extract. Browser file readers support text-based files, PDFs, DOCX, XLSX, PPTX, and supported text entries inside ZIP or skill packages, subject to scan limits.
- Browser protection doesn’t cover every website, native desktop app, or non-Chromium browser.
- Browser checks don’t provide OCR coverage for images or image-only scanned PDFs.
- The dedicated hidden-instructions check is PDF-specific. It isn’t a general hidden-text check for every Office format.
- Unreadable, unsupported, or partially scanned content must not be treated as fully checked.
A clean result means no risk was detected by the configured checks. It doesn’t guarantee that content is completely safe.
Cortex Guard helps reduce risk, but it doesn’t replace access controls, your organization’s security policies, or careful review.
Good security habits still matter
- Check where a file or skill came from before using it.
- Avoid sharing passwords, API keys, or confidential information unnecessarily.
- Treat unexpected instructions inside documents as suspicious.
- Don’t bypass a warning just to finish a task. Follow your organization’s policy.
- Ask your administrator if you’re unsure why content was blocked.
Pricing and activation
Cortex Guard is a separately priced add-on for enterprise plans. It isn’t included automatically in the standard TextCortex subscription.
Billing is based on purchased organization seats, including unassigned seats, not only seats currently assigned to users. Protection applies to eligible users holding covered seats.
Your administrator can review the organization’s seat count, current price, billing interval, subscription status, and total in the Cortex Guard settings. Please contact your account manager for details.