- Meta Red Team X found that an iPhone audio service let an app forge its identity and record after the user denied microphone permission. Apple has patched it.
- The missing orange indicator was confirmed only on an earlier iOS 27 beta build. Another researcher reproduced the bypass on iOS 26.6.
- Only iOS and iPadOS were tested. Check that devices run the August 17 releases, and note that iOS 27 release notes do not list the flaw.
A "no" that did not hold
Imagine a person who taps "Don't Allow" on a microphone prompt. They reasonably believe the matter is closed. According to research from Meta Red Team X, on affected systems it was not. An app could record anyway.
The idea worth taking from this flaw is simple. A permission prompt is what the user sees. The real decision happens deeper, in a check on who is asking. If that check can be fooled, the prompt is decoration.
What the researchers found
Two researchers documented the flaw. Nik Tsytsarkin, writing for Meta Red Team X, gave the full technical account. Mustafa Calap, a security engineer, had reported the same bug independently and released his own analysis on the same day, The Hacker News says.
The flaw is tracked as CVE-2026-65339. Tsytsarkin analysed two iOS 27 betas. Beta 4 was vulnerable and beta 6 contained the fix. On the earlier build, a sandboxed app recorded microphone audio with no permission grant and no special entitlements. The app did declare NSMicrophoneUsageDescription, the key that supplies the text for the permission prompt. Without it, the audio library stops the app.
On that earlier build, the orange microphone indicator stayed off. The system logs did not attribute the recording to the app. Tsytsarkin's demo video was filmed on a later build, and the dot is visible there.
Calap reproduced the bypass on iOS 26.6 and on an iPadOS 27 beta. His app recorded while its permission read denied, and no indicator appeared. The Hacker News says he does not state which of his two test builds that was. So the missing indicator is confirmed only on the earlier test build. The sources do not say whether shipping iPhones lost it.
In Tsytsarkin's test, the permission dialog showed only on the first launch. Tapping "Don't Allow" changed nothing, because recording went on regardless. The dialog did not return on later launches.
How the bypass worked
Apple's audio daemon, audiomxd, decides whether an app may record. It identifies the app with an audit token. That is a 32-byte value that describes a process.
The daemon exposed a setter that let the app supply its own token. The daemon copied those bytes in and then rechecked recording permission. Part of that recheck compared the token against the daemon's own identity. A match made the daemon treat the app as one of its own internal audio clients. It then allowed recording and skipped the call to TCC, Apple's system for managing app permissions.
So the app sent a token naming the daemon, and the daemon trusted it. The researcher notes that the recording was then attributed to the daemon, which explains the empty logs.
The app first needed the daemon's process ID and version number. Tsytsarkin shows that a sandboxed app could get both from the kernel-signed trailer on an error reply from the audio service. Guessing the version also worked. In his tests, trying values from 0 to 1023 turned up the right one in roughly 70 seconds.
An old mistake in a new place
Tsytsarkin places the bug in a known family. In 2014, Ian Beer found that OS X's networkd trusted an identity value taken from a message body. Apple fixed it as CVE-2014-4492. A 2020 post by Wojciech Reguła described a related failure.
The lesson is that the source of identity matters more than the check. Data the sender controls cannot prove who the sender is. The kernel's own record of the caller can.
Apple's fix follows that logic. The researcher reports that the patched build now checks the caller's real, kernel-supplied token. Only a caller holding a private entitlement may override the stored client identity. An app cannot grant itself that entitlement.
What the findings do not show
The Hacker News reports that neither Apple nor the two researchers say the flaw was used in attacks. It also reports that CISA scores the flaw 5.0 on the 10-point CVSS scale. According to the same outlet, the flaw was absent from CISA's known exploited vulnerabilities catalog at its most recent update.
An attacker would first need an app on the device. Calap installed his two test apps through Xcode, Apple's developer tool, so they skipped App Review. That means the tests cannot say whether Apple's review process would have caught such an app.
Calap found that recording had to start in the foreground. It then continued in the background and with the device locked. The bypass reached microphone capture through the audio service and nothing else. It did not grant other privacy permissions.
Neither researcher offers users a way to learn whether an app recorded them before the patch. Apple's own wording for the fix was brief: a logic issue, addressed with better checks. Tsytsarkin argues that it leaves out "the most striking part": recording after denial, with no indicator.
What to ask your team
Meta's timeline says the fix first appeared in iOS 27.0 beta 5 on August 10. The stable releases followed on August 17: iOS 26.6.1, iPadOS 26.6.1 and macOS Tahoe 26.6.2. The Hacker News says no workaround is described, so installing an update is the remedy. Ask how many managed devices still run earlier builds.
Scope matters here. The researchers tested iOS and iPadOS only. macOS has a listed fix, but no researcher reports testing it. Teams on iOS 27 should know the release notes do not list this flaw, so they cannot confirm their status from the notes. The fix shipped in beta 5, according to Meta's timeline.
For any app that handles sensitive conversations, ask the vendor how it decides who may access audio, files or cameras. The question to put to them is whether the identity comes from the platform or from the caller.
Last, check what your security team can see. On the earlier test build, neither the indicator nor the logs showed the recording. Controls that rely on those signals would not have helped.
A prompt asks permission. A check enforces it. This flaw is a case where the two came apart.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Meta Red Team X (Nik Tsytsarkin).





