BlackTree Security · Infrastructure · Automation · AI

BragJack Turned Extension Access into AI Agent Commands

The dangerous instruction did not arrive through a poisoned webpage. It came from a browser extension using the same permissions many users associate with ad blockers and content tools, then crossed into a far more powerful AI assistant.

Security researcher Gal Weizman calls the technique BragJack. The demonstrations required an attacker-controlled extension to be installed. A single proof-of-concept extension was adapted to attack Chrome Gemini Live, Perplexity Comet, Microsoft Edge Copilot Actions, Opera Neon and Claude in Chrome. The last is itself an extension, rather than a fifth standalone browser. Depending on the product and the capabilities exposed to its assistant, the demonstrated effects included local-file access, screenshots, browsing data, camera or microphone activation and authenticated actions. These were not identical capabilities across all five targets.

The extension forced the prompt through a trusted channel

BragJack is not ordinary prompt injection. The extension did not hide instructions in untrusted page content and hope the model followed them. It abused product-specific communication paths so the agent received attacker-authored commands from a context it treated as trusted. Weizman describes this as prompt forcing.

The research identifies the Chrome finding as CVE-2026-0628. CISA’s additional assessment assigns it a CVSS 3.1 score of 8.8; Google labels the WebView policy-enforcement flaw High severity. The related Edge race condition is tracked as CVE-2026-55945. Microsoft rates that information-disclosure issue Moderate, with a 4.2 base score. In its 22 September analysis, the Cloud Security Alliance reports that Perplexity, Opera and Anthropic remediated the demonstrated behaviours without comparable public CVE advisories in the sources it reviewed. That is a dated account of those demonstrations, not independent confirmation that every current agent build is free of related flaws.

Five products exposed different trust boundaries

Each product failed differently. In Chrome, request-rewriting rules reached the HTTPS content inside the privileged Gemini WebView even though other extension interactions were blocked. In Comet, a testing subdomain lacked the protection applied to the production agent domain. Opera allowed an extension to attach a content script to its own application domain. Those paths let the researcher impersonate the trusted component issuing commands to the assistant. The common problem is the boundary between ordinary extension capability and an agent that can act with the user’s authenticated context.

Edge exposed a race around a private tool-enablement action reached through its marketing page. The Edge and Claude in Chrome demonstrations also used debugger permission to generate clicks that their target interfaces accepted as user gestures. Describing every case as an ordinary extension silently taking over every assistant would lose those important prerequisites.

Update, 31 October 2026: the vendor records put the fixes in context

Google announced the Chrome fix on 6 January, in 143.0.7499.192 or .193 for Windows and macOS and 143.0.7499.192 for Linux. The published vulnerability description requires persuading a user to install a malicious extension. These are historical first-fix versions, rather than the browser versions to deploy now. Use the currently supported release; our Chrome October security review covers the newer update.

Microsoft’s Edge release notes date build 150.0.4078.48 to 2 July, while its security advisory is dated 3 July and identifies the same fixed build. The advisory describes a local attack by an authorised attacker that must win a race, with possible file-content disclosure and modification rather than loss of availability. Its scoring vector says no user interaction, while its explanatory FAQ describes two successive taps on deceptive or invisible forms. Both descriptions remain in the vendor record; the research’s debugger-generated gestures are relevant context, not a basis for rewriting Microsoft’s scoring.

Microsoft’s disclosure assessment said the flaw was neither publicly disclosed nor known to be exploited when that advisory was issued. That historical assessment does not establish the situation on 31 October. The BragJack research itself still does not establish a malicious campaign. For deployment, confirm the applicable current browser and agent updates rather than treating either July build or an acknowledgement of a proof of concept as a general assurance.

Six checks for managed browser extensions

  • Inventory extensions across managed browsers, including sideloaded and developer-mode installations.
  • Review requests for broad site access, request modification, debugger or scripting permissions as privileged access decisions.
  • Update Chrome, Edge and the affected agentic-browser products to builds containing the vendor remediations.
  • Separate experimental agent browsers from privileged administration, finance and sensitive data workflows.
  • Monitor unexpected AI-agent actions such as file reads, screenshot capture, camera or microphone access and navigation to unfamiliar destinations.
  • Require a fresh security review when an existing extension gains new permissions or changes ownership.

The research does not establish a malicious campaign using BragJack, and installing an attacker-controlled extension remains a prerequisite. That prerequisite is not trivial, but browser extensions are already a mature supply-chain and social-engineering target. The significance lies in what the same foothold can now reach.

The fixes for CVE-2026-0628 and CVE-2026-55945 close two implementations. BragJack’s wider lesson is that an assistant able to browse, read and act should not inherit trust merely because a message appears to originate inside the browser.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *