The counters above are scanner output, not vulnerabilities. They are published for audit transparency, and they are expected to be large for four reasons.
The same issue is reported many times. Thirty-six agents review the same code independently, each with a different specialty, so one underlying flaw surfaces once per agent that notices it. In this scan a single unsanitised-HTML pattern in the DevTools inspector was reported three separate times as three findings. It is one issue with one fix.
The scanners are built to over-report. Most raw volume comes from deterministic pattern matching, which is deliberately high-recall and low-precision. It is cheap and exhaustive, and its job is to miss nothing so that the judging stage has everything to work from. Precision is added later, on purpose.
Agents cannot see the environment the code runs in. A reviewer sees source, not runtime. It cannot know that a file is test scaffolding, that a path is unreachable in production, or that a platform mitigation already neutralises the issue. That is not a defect in the agent -- it is the reason a judging stage exists.
One architectural problem fragments into many findings. A missing validation layer or an unbounded deserialisation pattern is reported at every site where it appears. Those are not separate bugs. They are one design decision with many symptoms, requiring a single fix.
Consensus merges agents that found the same thing. Cross-chunk linking runs before the false-positive filter deliberately, because a finding that looks harmless in one file can become serious when connected to two others -- filtering first would discard real threats. The false-positive filter then removes test code, detection code and documentation. Cascade and architectural passes collapse repeated symptoms back to their structural cause.
Only then does the judge model read what survives, with that cross-chunk and architectural context attached, and decide what is genuinely exploitable, what severity it actually warrants, and what to discard. The raw number measures how hard the scanners looked. The confirmed number is the finding.
setInnerHTML, setOuterHTML, and insertAdjacentHTML in walker.js accept user-controlled HTML without sanitization. These are DevTools server-side methods that can be called by privileged actors. If an attacker can control the HTML content passed to these methods (e.g., via a compromised extension or malicious inspected page), they can inject arbitrary scripts that execute in the DevTools UI context. This is a high-severity XSS vector. PoC testing confirmed the injection path exists but Firefox process isolation prevents escalation to chrome-privileged context.generateGmpJson.py enable SSRF. An attacker who can control the GMP JSON configuration can make the browser fetch arbitrary URLs, potentially accessing internal services.socket.js via unvalidated host parameter. An attacker can make the DevTools server connect to arbitrary WebSocket endpoints.GetFileOrDirectoryTask.cpp allows accessing files outside the intended directory.interventions.js restartListener. An attacker can cause the browser to hang by providing many URLs.sprintf.js regex patterns. An attacker can provide format strings that cause exponential backtracking.nsContentSecurityUtils.cpp regex patterns.StorageAccessPermissionRequest.cpp. An attacker who can observe timing can predict the simulated delay, potentially bypassing storage access restrictions.OggWriter.cpp.MediaParent.cpp.ClearKeySessionManager.cpp -- inserting Key instead of KeyId into mKeyIds set. This can cause incorrect key management in EME.nsIWindowMediator enumeration from listener. An attacker can trigger a deadlock by opening many windows.ContentEventHandler.cpp GetTextEditor. An attacker can trigger a crash by manipulating the DOM.javascript: protocol validation in checkProtocol function. An attacker can bypass URL validation.ext-url-overrides.js. An attacker can make the browser navigate to arbitrary URLs.PrincipalUtils.cpp. An attacker can bypass security checks.nsXMLFragmentContentSink.cpp. An attacker can perform XXE attacks.CreateProcess failure in LauncherProcessWin.cpp. An attacker can exhaust system resources.DecoderTemplate.cpp. An attacker can cause inconsistent state.GetData.setInnerHTML, setOuterHTML, and insertAdjacentHTML calls in walker.js) can execute arbitrary JavaScript in the DevTools UI context. Firefox process isolation prevents escalation to a chrome-privileged context in current builds. Rated High.setInnerHTML, setOuterHTML, and insertAdjacentHTML.The high-severity issues that require immediate attention are: