top of page

The Spy Who Listened After Being Denied: How an iPhone Vulnerability Turned a Trusted Audio Service Into an Eavesdropping Opportunity

hace 1 día
11 min de lectura

The building was designed to protect its owner’s privacy.

Every room had secure doors. Every conversation was confidential. Every listening device required explicit authorization.

At the center of the building stood a surveillance control room.

Its operators managed the microphones installed throughout the property.

But there was one fundamental rule.

Nobody could activate a microphone without the owner’s permission.

One evening, an unfamiliar visitor approached the security desk.

He requested access to the building’s listening system.

The guard contacted the owner.

“Do you authorize this visitor to listen to your conversations?”

The answer was immediate.

“No.”

The guard rejected the request.

The visitor walked away.

But he wasn’t an ordinary visitor.

He was a spy.

And he had discovered something unusual about the building’s security procedures.

The guards carefully checked the identities of ordinary visitors.

But when the head of surveillance requested access, they assumed the request was legitimate.

The spy realized he didn’t need to persuade the owner.

He needed to convince the system that he was the head of surveillance.

So he returned with a different identity.

This time, the guard opened the door.

The microphones activated.

And the owner continued speaking, unaware that his explicit refusal had been bypassed.

The spy hadn’t defeated the owner’s decision. He had found a way to make the security system ignore it.

This is the central idea behind CVE-2026-65339, a vulnerability in Apple’s audio service disclosed in technical detail in November 2026.

Researchers demonstrated that an installed application could record audio on affected iPhones and iPads even when microphone permission had been denied.

In some tests, the recording occurred without the familiar orange microphone indicator appearing.

The flaw has since been patched.

Neither Apple nor the researchers reported evidence of exploitation in real-world attacks.

Nevertheless, the discovery exposed a fundamental problem in how a trusted system service verified the identity of applications requesting access to sensitive hardware.


Phase 1 — The Owner Controls the Listening Room


The building’s privacy policy was straightforward.

Visitors could request access to the listening room.

But the owner had the final say.

If he refused, the guards were supposed to prevent access.

Apple implements a similar principle through its Transparency, Consent, and Control (TCC) framework.

TCC manages application access to sensitive resources, including the microphone.

When an application requests microphone access for the first time, iOS presents a permission prompt.

The user can approve or deny the request.

Under normal circumstances, an application that receives a denial should not be able to record audio through that protected interface.

The decision belongs to the user.

Not the application.

Not the developer.

And certainly not an unauthorized third party.

The owner had established the rules. The security system was responsible for enforcing them.


Phase 2 — The Spy Receives an Explicit Refusal


The researchers’ experiments began with applications that requested microphone access.

The user was shown the standard permission prompt.

The user selected “Don’t Allow.”

That should have ended the matter.

But the test applications were still able to record.

Nik Tsytsarkin, a researcher at Meta Red Team X, demonstrated that denying microphone permission did not prevent recording in the affected builds.

The prompt appeared when the application first opened.

After the user rejected access, recording could still proceed.

The prompt did not reappear during subsequent launches in his tests.

Mustafa Calap independently reported the same underlying vulnerability.

His test application also recorded audio while its microphone permission remained denied.

The result was particularly concerning because the application did not need to possess special private entitlements.

The demonstrated bypass involved ordinary sandboxed test applications.

However, those applications were installed directly through Xcode.

They were not submitted through Apple’s App Review process.

Consequently, the research did not establish whether such applications would have passed App Store review.

The spy had requested permission. The owner had refused. But the refusal did not prevent the listening system from being activated.


Phase 3 — The Orange Warning Light


The building had another protective mechanism.

Whenever a microphone was activated, an orange warning light was supposed to illuminate.

Even if the owner wasn’t watching the security desk, he could see when someone was listening.

On iPhones, the orange microphone indicator serves a comparable purpose.

It provides a visible signal when an application is using the microphone.

But the researchers observed something unexpected.

In some affected builds, audio recording occurred without the indicator appearing.

Tsytsarkin reported that the orange indicator remained absent in an earlier test build.

His later demonstration video showed the indicator appearing despite microphone permission remaining denied.

Calap reported tests in which neither Dynamic Island nor Control Center displayed the microphone indicator.

These distinctions are important.

The vulnerability did not universally suppress the indicator in every tested build.

Its most consistent demonstrated effect was the bypass of the microphone permission decision.

The absence of the orange indicator was an additional behavior observed in particular tests.

In some cases, the spy could activate the listening room without even switching on the warning light.


Phase 4 — The Spy Studies the Security Chief


The spy began investigating the building’s surveillance infrastructure.

Ordinary visitors were required to identify themselves before accessing the listening system.

But the head of surveillance operated under a different set of assumptions.

Certain internal operations were trusted because they originated from the surveillance service itself.

The spy discovered that the guards could be fooled about who was making a request.

This corresponds to the technical weakness in Apple’s audio service, audiomxd.

The service handles audio-related operations and interacts with the system’s permission enforcement mechanisms.

When an application requests microphone access, the system needs to establish which process is responsible for the request.

That identity is represented through an audit token.

An audit token identifies a process and provides information that can be used to make security decisions.

The problem was that the vulnerable audio service accepted an application-supplied token as the basis for determining the requesting identity.

The application could provide a token that identified the audio service itself.

Instead of recognizing the actual external caller, the service treated the request as though it originated internally.

The spy didn’t need to become the security chief. He only needed the guards to accept a false identity claiming that he was.


Phase 5 — The False Identity Opens the Door


The exploit depended on a trust-boundary failure.

An unprivileged application should not be able to decide which identity a privileged service uses when evaluating a sensitive request.

That identity should come from a trusted system source.

But the vulnerable implementation allowed the requesting application to influence the identity presented to audiomxd.

By supplying a token representing the audio service, the application could make the service believe it was handling its own internal audio operation.

The service then followed a privileged internal path.

That path did not require the normal TCC microphone permission check.

The result was unauthorized microphone access.

This was not a conventional attack in which the user was tricked into tapping “Allow.”

Nor did it involve the application legitimately receiving microphone permission and later misusing it.

The researchers demonstrated that the permission could remain explicitly denied.

The guards didn’t disregard the owner’s refusal because they had been ordered to. They disregarded it because they believed the request came from someone who didn’t need the owner’s permission.


Phase 6 — The Kernel Supplies the Identity


One of the most interesting details concerns where the forged identity originated.

The researcher explained that the application could obtain the information needed to impersonate the audio service from the operating system’s kernel.

The weakness was not that the kernel had been compromised.

Rather, the problem was how the audio service handled the identity information it received.

A service making a security-sensitive decision should rely on the authenticated identity of its actual caller.

Instead, the vulnerable implementation accepted an identity that the application could supply.

That distinction is central to understanding the flaw.

The system had a trusted mechanism for identifying processes.

But the service was not consistently using that mechanism to determine who was really making the request.

In the building, the security desk possessed an official identity register.

But instead of checking the visitor against that register, the guard accepted an identity document provided by the visitor.

The security system had the information needed to identify the spy. The weakness was that it trusted the identity the spy presented instead.


Phase 7 — The Spy Keeps Listening After the Door Closes


Calap’s testing revealed another important behavior.

The recording needed to begin while the application was in the foreground.

But once recording had started, it could continue after the application moved into the background.

It could also continue after the device was locked.

The background-recording demonstration used an application configured with the audio background mode.

This did not establish that every application could begin recording at any moment while suspended.

The initial activation conditions still mattered.

But the ability to continue recording after the application was no longer visible increased the privacy implications.

In the building, the spy entered the listening room while the owner was nearby.

Then the owner left.

The lights went out.

The doors closed.

But the microphones continued recording.

The spy didn’t necessarily need to remain visible to continue listening.


Phase 8 — What the Spy Could and Couldn’t Access


The vulnerability was serious because it undermined an explicit privacy control.

However, its scope must be understood accurately.

The researchers demonstrated unauthorized microphone recording.

They did not demonstrate that the same vulnerability granted access to every protected resource on the device.

The bypass did not automatically provide camera access.

It did not establish unrestricted access to photographs, contacts or messages.

It was not described as a remote, zero-click compromise of an untouched iPhone.

An application had to be installed on the affected device.

And the demonstrated technique targeted microphone capture through the vulnerable audio service.

These limitations do not eliminate the privacy risk.

Conversations may contain personal information, business discussions, credentials spoken aloud or other sensitive material.

An application recording audio without authorization could expose that information.

But the published research did not demonstrate that a real attacker had collected or exfiltrated such recordings.

The spy had found a way into the listening room. That did not mean he possessed the keys to every other room in the building.


Phase 9 — Two Investigators Discover the Same Weakness


The vulnerability was independently reported by two researchers.

Nik Tsytsarkin of Meta Red Team X published technical details describing the failure in Apple’s audio service.

Security engineer Mustafa Calap separately investigated and reported the same underlying problem.

Calap reproduced the bypass on iOS 26.6 and an iPadOS 27 beta.

Tsytsarkin’s analysis compared two iOS 27 test builds.

Their findings reinforced the conclusion that the problem involved the audio service’s handling of caller identity.

They also helped clarify differences in the behavior of the microphone indicator across test environments.

Neither researcher reported evidence that malicious actors had exploited the vulnerability in the wild.

Their work demonstrated a security weakness through controlled testing.

Two independent investigators examined the building’s security system and discovered that the same false identity could defeat its microphone-access controls.


Phase 10 — The Security Chief Learns to Verify Real Identities


Apple corrected the underlying issue by changing how the audio service validates requests.

Instead of trusting the identity supplied by an external application, the patched service checks the actual caller identity provided by the kernel.

This closes the impersonation path demonstrated in the research.

External callers wishing to override the client identity now require a restricted private entitlement:

com.apple.private.audio.client-audit-token-override

An ordinary application cannot simply grant itself that entitlement.

The change restores the distinction between a genuine internal operation and a request made by an outside application.

In the building, the guards received a new instruction.

No visitor could identify themselves as the head of surveillance merely by presenting a document.

The guard had to verify the visitor’s actual identity against the official register.

The solution wasn’t to ask the owner to say “No” more forcefully. It was to ensure that the security system respected that answer regardless of the identity an intruder claimed.


Phase 11 — The Repairs Across Apple’s Ecosystem


The flaw is tracked as CVE-2026-65339.

CISA assigned it a CVSS severity score of 5.0 out of 10.

Apple released fixes across several operating systems.

Platform

Fixed release

iPhone 11 and later

iOS 26.6.1

Compatible newer iPads

iPadOS 26.6.1

iPhone XS, XS Max and XR

iOS 18.7.11

iPad 7th generation

iPadOS 18.7.11

macOS Tahoe

26.6.2

macOS Sequoia

15.8

Apple TV

tvOS 27

Apple Watch

watchOS 27

Apple Vision Pro

visionOS 27

The first fixes arrived on August 17, 2026.

Additional fixes for supported older devices followed in October.

Although Apple listed fixes for several operating systems, the researchers’ published demonstrations focused on iOS and iPadOS.

They did not report equivalent tests on macOS, tvOS, watchOS or visionOS.

The Meta research timeline also indicated that the fix was included in iOS 27.0 beta 5.

As of November 17, the vulnerability was not listed in CISA’s Known Exploited Vulnerabilities catalog.

The vulnerability had been discovered and the locks repaired. But users still needed to install the updated security mechanisms.


Phase 12 — How to Protect the Listening Room


The researchers did not identify a user-accessible setting that reliably prevented the bypass on vulnerable systems.

That means the primary corrective action is to install the relevant Apple security update.

Organizations and individual users should take several precautions.

Update affected devices. Install the latest compatible iOS or iPadOS release containing the fix.

Maintain device inventories. Organizations managing iPhones and iPads should identify devices that have not received the relevant security updates.

Control application installation. Use trusted distribution channels and organizational application-management policies where appropriate.

Review unnecessary applications. Remove software that is no longer needed or comes from untrusted sources.

Maintain mobile security visibility. Enterprise environments should use appropriate device-management and compliance controls to identify outdated devices.

Avoid relying exclusively on permission prompts. Privacy controls are essential, but their correct enforcement also depends on the security of the underlying operating system.

Understand the indicator’s limitations. The absence of an orange microphone indicator on an affected build was not conclusive proof that recording was impossible.

Monitor vendor advisories. Follow Apple’s security guidance for future changes and additional fixes.

These measures should not be interpreted as evidence that a particular installed application has exploited the vulnerability.

The published research did not provide a method for users to determine retrospectively whether an application had recorded audio before the patch.

The best defense wasn’t to watch the warning light more carefully. It was to repair the mechanism that determined who could activate the microphones.


Phase 13 — The Larger Problem of Trusted Identities


The implications extend beyond one audio service.

Modern operating systems depend on complex interactions between applications, privileged services and security frameworks.

Those components must communicate.

They must identify one another.

And they must make decisions about which operations are authorized.

When a privileged service accepts identity information from a less-trusted component without properly validating it, a trust boundary can collapse.

That is what made this vulnerability noteworthy.

The microphone permission system existed.

The user made a clear decision.

The operating system had mechanisms for identifying processes.

But the audio service trusted the wrong source when determining the caller’s identity.

This allowed a lower-trust application to obtain an operation that should have been restricted.

The same general security principle applies throughout computing:

A system must not allow the party requesting access to unilaterally decide which identity the system should trust.

The discovery illustrates why authorization depends not only on the rules themselves, but also on reliable authentication of the entity to which those rules apply.


The Final Scene — The Spy Removes His Disguise


The owner returned to the security desk.

He had refused microphone access.

He remembered the warning.

He remembered selecting “Don’t Allow.”

And he believed the listening room had remained closed.

But the investigators showed him what had happened.

A spy had entered the building with limited access.

He had approached the surveillance system.

He had presented a false identity.

And the guards had accepted him as the head of surveillance.

In some tests, the warning light had remained dark.

In others, it had illuminated, but the owner’s denial had still been bypassed.

The flaw had allowed recording without legitimate microphone permission.

There was no evidence that criminals had used the technique in real-world attacks.

And Apple had already corrected the underlying weakness.

But the discovery revealed something deeply important about digital privacy.

A permission prompt is only as reliable as the system enforcing it.

The owner can refuse.

The application can remain unauthorized.

The interface can show that permission has been denied.

Yet if a privileged service accepts a forged identity, those protections may be circumvented.

The spy removed his disguise.

The guards finally recognized him.

And the building’s security procedures were rewritten to ensure that the same deception would no longer work.

The most disturbing part wasn’t that someone could listen without permission.

It was that the owner had explicitly said no, and the system had still allowed the spy to listen.


 
 
 

Comentarios


bottom of page