ShieldBreak: Microsoft Defender Bypass Analysis

A technical advisory review covering the reported Defender bypass chain, defensive checks and security engineering lessons.

What actually broke

The starting point is CVE-2026-50656, a race condition and link-resolution flaw inside mpengine.dll — the scanning engine Defender itself runs on — that a local attacker could use to escalate to NT AUTHORITY\SYSTEM. Microsoft shipped a fix for it in July 2026, in engine build 1.1.26060.3008, and for about a month that looked like the end of it.

On 11 August 2026, a researcher publishing as "Chaotic Eclipse" released ShieldBreak: not a new hole in Defender, but a different route into the same underlying race the July patch was supposed to have closed. Independent write-ups describe a consistent, repeatable success rate against fully patched Windows 11 25H2 and Windows Server 2025 hosts. Microsoft has confirmed it is investigating, rated exploitation "more likely," and had not shipped a fix for the bypass itself as of this writing.

How the bypass actually works

The chain leans on parts of Windows that have nothing to do with malware scanning on the surface. The attacker registers a rogue cloud storage provider and attaches it to a crafted placeholder file, then uses Object Manager symlinks together with CLFS (the Common Log File System) to steer Defender's own cloud-hydration scan toward a file that should not exist on a stock install: system32\phoneinfo.dll.

While Defender holds that file locked through the Cloud Filter API mid-scan, the attacker swaps in a malicious phoneinfo.dll. The trigger for actually running it is almost mundane: Windows Error Reporting's own QueueReporting scheduled task later launches wermgr.exe with elevated rights, wermgr.exe loads the planted DLL as if it were the real system library, and the attacker ends up with a conhost.exe shell running as SYSTEM.

One caveat worth being precise about, because a lot of the coverage glosses over it: this is a local privilege-escalation chain, not a remote way in. It assumes the attacker already has ordinary code execution on the box — through phishing, a different exploit, or a foothold from an earlier stage — and turns that into full SYSTEM control. That is exactly why ransomware operators and red teamers both care about LPE: it is almost always the step between "we got a user's laptop" and "we own the domain."

The part we found genuinely interesting is what it says about the July patch, not just about Defender. The fix closed the specific path the original researcher demonstrated. It did not remove the race condition underneath it. ShieldBreak is what happens when a second researcher goes looking for a different door into the same room — a pattern worth assuming on any high-value privilege-escalation fix, not just this one.

What we are checking for on Windows engagements this quarter

With no patch for the bypass itself, this is a detection-and-containment problem until Microsoft ships one, not a "patch and move on" problem.

Inventory comes first: find out where Defender is actually the registered, active antivirus. That is the precondition for this exact chain — endpoints running a different primary AV are not exposed to this specific technique, whatever else they need patched.

Treat the interim mitigations as exactly that, interim. Locking a zero-byte placeholder at system32\phoneinfo.dll blocks the specific substitution step this chain relies on, but it is a compensating control, not a fix — it should go through a small pilot ring with a rollback plan before it touches production, since forcing a file to exist where the OS does not expect one can have side effects. Confirming Defender tamper protection is actually on, and tightening ASR rules and application allowlisting, raises the cost of getting the local code execution this whole chain depends on in the first place.

On the detection side: watch for wermgr.exe spawning child processes with upload-style arguments, flag non-SYSTEM processes writing into the WER ReportQueue path, and treat a conhost.exe with an unusual parent chain as worth a second look. None of these are exotic signals to add to a watchlist.

And because privilege escalation is a normal phase of an authorised penetration test, this is now on our list to check for on any Windows-based internal or red-team engagement scoped this quarter — if a technique like this works during an authorised test, it works for whoever gets there first without permission.

Explore related information