A BragJack browser agent attack disclosed by security researcher Gal Weizman of Forever Security can seize control of AI assistants built into popular browsers using a single malicious extension, without any interaction from the victim. Published on 16 September 2026 and detailed by TechTimes, the proof-of-concept ran against five targets: Google Chrome’s Gemini Live, Perplexity Comet, Microsoft Edge, Opera Neon, and Anthropic’s Claude in Chrome.

The research produced two CVEs and earned more than $20,000 in bug bounties from the five vendors, with individual payouts ranging from $600 to $7,000. Both Google and Microsoft have since patched the vulnerabilities assigned to them.

How the BragJack Browser Agent Attack Works

Weizman frames AI browser assistants as having a ‘brain’ and a ‘body.’ The model decides what to do; a privileged browser component carries out the action, whether that is reading tabs, taking screenshots, accessing files, or interacting with websites on the user’s behalf. The weakness is that browser extensions can manipulate the web traffic and pages those privileged components trust, and BragJack exploits exactly that gap.

The same extension was used across all five targets, relying on Chromium’s declarativeNetRequest (DNR) functionality. DNR lets extensions modify how network requests are handled, including rewriting response headers and redirecting resources. According to Gblock, the extension requests DNR permissions and access to all HTTPS hosts to inject code into whichever website each AI assistant trusts. The Edge and Claude exploits required an additional debugger permission on top of that.

In the Chrome attack, extensions cannot directly touch the privileged chrome://glic component or inject scripts into Google’s Gemini site. Weizman found, however, that DNR rules could still intercept requests made by the embedded Gemini web app. By weakening security headers and redirecting a JavaScript resource, he executed code inside the Gemini context and communicated directly with Chrome’s privileged AI component. The resulting access could read local files, reach web content, take screenshots, and potentially reach the browser’s camera and microphone. Chrome assigned the finding CVE-2026-0628, paid a $7,000 bounty, and issued a fix in Chrome 143.0.7499.192.

Agentic Browsers and the Stakes of Prompt Forcing

The attacks against agentic browsers go further, because those agents can act on websites rather than merely read them. For Perplexity Comet, Weizman found the browser’s built-in agent extension trusted several Perplexity domains, including a testing domain that did not receive the same protections as the main perplexity.ai site. Removing a redirect to that domain with DNR let him load it and inject a content script able to talk directly to the built-in agent. The resulting access covered browsing history, screenshots, local files, and the ability to issue instructions to the agent. He demonstrated forcing Comet to visit Perplexity, summarise the victim’s emails, and send the results to another address. Perplexity paid a $7,000 bounty, but Gblock reports no CVE was assigned and no fix date for the exact method has been confirmed.

Microsoft Edge presented a different problem. Microsoft had split its agent into ‘Think’ and ‘Do’ modes to prevent it from taking arbitrary instructions and actions simultaneously. Weizman found a race condition that briefly disables the restriction while a prompt is forced, then re-enables the action capability before the agent checks its state. Microsoft assigned CVE-2026-55945 to the race condition and shipped a fix in Edge 150.0.4078.48 on 2 July 2026. Similar flaws were demonstrated against Opera Neon and Claude in Chrome, with the latter being a browser extension itself rather than a native browser feature.

Weizman calls the underlying technique Prompt Forcing. Unlike conventional prompt injection, where an attacker tries to slip malicious instructions into content the AI is already reading, Prompt Forcing hands the agent an entire prompt and follow-up instructions directly. The agent then translates those instructions into legitimate browser actions using its existing privileges. The practical consequence for defences is that the final action is not carried out by conventional malicious code: legitimate software is being directed to perform the attack.

The BragJack browser agent attack points to a growing challenge as browsers add more capable AI agents. A compromised extension that would traditionally see only web content can, in some designs, become a path into software that reads local files, browsing history, and acts on websites on the user’s behalf. Weizman advises keeping browsers fully updated, removing extensions that are unfamiliar or no longer needed, and treating broad ‘read and change all your data on all websites’ permission requests with caution. He has published a full technical breakdown covering all five attacks alongside the research disclosure.

Share.

Software engineer and video game uber-nerd.

Comments are closed.

Exit mobile version