Vigil Guard 1.8.6 Extension Verified in Live Sessions, Vulnerability Review Closed

Two parts of AI protection can become outdated even when nothing changes on the customer side. The first is the integration with AI provider applications. ChatGPT, Claude, Gemini, Copilot and the other supported services evolve independently of us, sometimes through a substantial rebuild, sometimes through a change that is barely visible to the user, and either way it can affect how the browser extension interacts with the application. The second is the set of third party components used across the platform, where our own code may remain unchanged while a new vulnerability is disclosed in a library, base image or another dependency. In both cases, protection that was correct on release day does not stay correct by itself.

Version 1.8.6 is the result of work across both of these areas. The update remains compatible with existing integrations and security policies, and requires no configuration changes or data migration.

The browser extension is one of the main control points for Shadow AI. This is where we see an employee prompt before it is sent, where the policy decision is made in real time, and where the event that later appears in the audit trail is created. The complication is that the extension operates inside applications we do not control. A provider may change the way a form works, modify an interface element or alter how a conversation is handled, and a difference that is almost invisible to the user can still matter for a mechanism that has to identify and protect the prompt correctly.

We therefore do not treat provider side changes as exceptional failures to be addressed only after a customer reports a problem. They are part of normal product maintenance. Our internal controls monitor the behaviour of supported services, and when one of them detects a change we do not test only the service that triggered the alert, we run the checks across the full set of supported applications. If the change requires an adjustment, the extension is updated accordingly, and we run the same cycle periodically and on customer request.

For version 1.8.6, the audit covered ChatGPT, Claude, Gemini, Microsoft Copilot for consumers and Microsoft 365, Perplexity, DeepSeek, Mistral, Meta AI, Poe, Z.ai and Bielik. Every service was tested in a real, authenticated session, and for each one we verified all three decision paths: ALLOW, SANITIZE and BLOCK. We also checked that protected content was no longer visible anywhere after returning to the conversation, and that a single action generated exactly one event in the audit trail. A list of supported services on its own does not tell you much, it has value only when every item on that list is backed by this level of testing.

The second major part of version 1.8.6 concerns third party components. Our rule is straightforward: we patch known vulnerabilities in the components we use, including cases where there is no known exploitation path in our architecture. We do not spend days trying to prove that a given vulnerability is definitely unreachable in our code, because establishing whether a specific issue can be exploited in a particular part of the system can take more time and effort than updating the affected library.

In 1.8.6 we completed a multiweek programme covering third party components across the entire platform. As part of that work we also reduced the contents of the images we ship, so they now carry only what is required in production, and restricted automation processes to the minimum permissions they need to operate.

There are, of course, cases where a fix is not yet available from the provider of the underlying system. When that happens, the exception is documented and given an expiry date. We do not leave such exceptions in the system indefinitely.

All 21 platform components were built and scanned for vulnerabilities from the exact version of the code that went into the release, not from an earlier build or an image created outside the release process. The scans apply to the same artefacts that were actually published.

Every released image is cryptographically signed, and the component signature list is signed as well. Customers can verify that list independently and confirm for themselves exactly what they are running.

Version 1.8.6 also includes several changes outside the two main areas of the release. One of them concerns dormant accounts, which are awkward from a security perspective precisely because nothing usually happens on them for long periods of time. If nobody normally uses an account, it is easier to miss the moment when activity appears for a completely different reason. Administrators can now see dormant accounts directly in user management and handle them from the same place.

We also fixed request limit behaviour, so the configured threshold is enforced correctly under concurrent traffic. The audit trail records events without silent gaps, and the viewer role is restricted to read only access.

Sensitive data detection has been expanded as well. Alongside Polish identifiers such as PESEL, NIP and REGON, the platform can now detect IBAN account numbers and several additional national identifiers, depending on configuration. This matters particularly in deployments where traffic is not limited to Poland.

Vigil Guard Enterprise 1.8.6 is available to all customers with active support. The installer detects an existing deployment and performs the upgrade while preserving both configuration and data. We recommend prioritising the update in environments where the browser extension is the primary control point for Shadow AI, and wherever security policy requires third party components to remain current.