Does WhatsApp Encrypt Backups? What "Encrypted" Actually Covers

You probably assume your WhatsApp conversations are private. End-to-end encrypted, the app tells you so in every chat. But does WhatsApp encrypt backups the same way it encrypts messages in transit? The answer is: not by default, and the gap between what most users believe and what the system actually protects is significant. Messages, backups, and metadata each live under different encryption regimes, with different defaults, different threat models, and different exposure surfaces. This piece breaks down all three.

Key Takeaways

• WhatsApp messages use the Signal Protocol for end-to-end encryption in transit between devices. Backups stored on cloud services are not end-to-end encrypted by default; you must enable that yourself.
• Even with encrypted backups turned on, WhatsApp still collects and retains a wide range of metadata: who you talk to, when, how often, your device info, group memberships, online status, and more. None of that is covered by E2EE.
• Messages sent to Meta AI inside WhatsApp are processed on Meta's servers and fall outside the standard E2EE model entirely.
• A 2025 vulnerability in WhatsApp's contact-discovery system exposed metadata for roughly 3.5 billion accounts, demonstrating that the attack surface is usually at the edges of an encrypted system, not in the cryptography itself.
• Asking "is it encrypted?" is the wrong question. The right question is a three-part audit: what happens to content, what happens to storage/backups, and what happens to metadata.

What Does WhatsApp's End-to-End Encryption Actually Cover?

It covers the content of messages, voice calls, video calls, photos, and documents while they travel between your device and the recipient's device. The Signal Protocol handles this. Neither WhatsApp nor Meta can read message content in transit. This part works as advertised and has held up well under scrutiny.

The problem is scope. "End-to-end encrypted" describes a narrow pipe: content, in motion, between two endpoints. Everything outside that pipe, including where that content gets stored after arrival and everything the system knows about the content without reading it, falls under different rules. WhatsApp's own documentation acknowledges that E2EE does not equal total privacy, though most users never register the distinction.

Are WhatsApp Cloud Backups Encrypted by Default?

No. By default, chat backups stored on iCloud or a cloud drive are not end-to-end encrypted. They sit on third-party infrastructure protected by whatever encryption the cloud provider applies at rest, which means the cloud provider (and anyone who compromises or compels the cloud provider) can access them. Users must manually enable encrypted backups for E2EE protection to extend to stored conversations.

This has been true since WhatsApp introduced cloud backups. The company added optional end-to-end encrypted backups later, but the timing gap was years long, and the default never changed. Most users who enabled cloud backup at some point are running unencrypted backups right now without knowing it.

How Do Encrypted Backups Work When You Turn Them On?

When you opt in, WhatsApp encrypts the backup with a key you control. You can choose a custom password or a 64-digit encryption key. In late 2025, WhatsApp added passkey support, letting you secure backups with a fingerprint or face scan instead of memorizing a password. The key itself is stored in an HSM-based Backup Key Vault, housed in tamper-resistant hardware that Meta says is inaccessible to Meta, cloud providers, or third parties.

In May 2026, Meta rolled out over-the-air key distribution and committed to publishing proof of secure hardware deployments for this vault infrastructure. That is a meaningful step. The architecture is sound on paper. But the default is still off. A feature that protects you only when you find and flip a buried toggle is, for most of the user base, a feature that does not exist.

What Metadata Does WhatsApp Collect Outside of E2EE?

A lot. WhatsApp collects account information, device and network details, usage logs, hashed contact lists, online and last-seen status, group membership, and call and media metadata. None of this is message content. All of it falls outside end-to-end encryption.

The hashed contact list is particularly interesting. WhatsApp uploads your phone's contact list (hashed, but still) to build a social graph. Who you know, who they know, who is in which group. This is structural data about your relationships, and it is available to Meta regardless of whether your messages are encrypted.

Online status and last-seen timestamps reveal behavioral patterns: when you wake up, when you go to sleep, how frequently you check your phone, whether two people are online at the same time. Call metadata reveals who called whom, for how long, and when. None of this requires reading a single message.

Why Does Metadata Matter If Message Content Is Protected?

Because metadata is often more useful than content for building a profile. A well-known formulation from intelligence work: "We kill people based on metadata." That is an extreme case, but the principle scales down. Your social graph, your communication patterns, your device fingerprint, your location signals embedded in network data, these compose a detailed picture of your life that does not require anyone to decrypt a single message.

Researchers at the University of Vienna and SBA Research put it directly: end-to-end encryption protects the content of messages, but not necessarily the associated metadata. The two are different categories of information, governed by different protections, collected under different policies.

What Did the 2025 Contact-Discovery Vulnerability Reveal?

In November 2025, University of Vienna and SBA Research academics disclosed that a vulnerability in WhatsApp's contact-discovery feature could enumerate approximately 3.5 billion accounts. The exposed data included phone numbers, public encryption keys, timestamps, and (if set to public) profile photos and "about" text. From this, researchers could infer device OS, account age, and linked-device count.

Note what was not exposed: message content. The Signal Protocol held. The cryptography was fine. The leak was entirely at the metadata and discovery layer, which is exactly the pattern that repeats across encrypted systems. The core encryption works. The edges leak. Meta has since addressed and mitigated the issue, but it remains a concrete example of how an encrypted system's real attack surface is rarely the encryption itself.

We see this pattern in our own work building Selina. When you run an encrypted-at-rest system, the database is not your source of truth for debugging. A content column can read as empty and still decrypt to valid data. You learn quickly that you have to debug through the application's own decrypt path on the specific rows under investigation, not by reading the raw column. The point is that encryption changes where complexity lives, and the edges of the encrypted boundary are where things actually break.

What Happens When You Message Meta AI Inside WhatsApp?

It leaves the E2EE model entirely. Messages sent to Meta AI inside WhatsApp are processed on Meta's servers to generate responses and maintain conversational context. This is a fundamentally different architecture from person-to-person E2EE. Users may share sensitive health, financial, or personal details in these conversations without clarity on retention or use.

Meta recognized this tension. In May 2026, they announced Incognito Chat with Meta AI, a mode where conversations are processed inside a Trusted Execution Environment so the provider architecturally cannot read the exchange. This is separate infrastructure from person-to-person E2EE. It is not an extension of the existing encryption; it is a new system built to address the fact that bolting AI features onto an E2EE app breaks the original privacy model.

This is a real architectural challenge, not a marketing problem. Any AI assistant that needs to understand your messages in order to respond has to, by definition, process them. The question is where that processing happens, who can observe it, and what persists afterward. TEEs are one credible mitigation. On-device inference is another, though it constrains model capability. Ephemeral processing with no persistence is a third. What matters is that you can verify the architecture, not just the policy promise.

How Does This Affect the Encryption Boundary?

It redraws it. When WhatsApp was a messaging app, the encryption boundary was clear: content encrypted between sender and receiver, metadata collected by WhatsApp. Adding an AI assistant creates a third party inside the conversation. That third party needs access to plaintext to function. The boundary now has a controlled opening, and the question becomes whether that opening is genuinely constrained or just policy-constrained.

Meta has clarified that while person-to-person messages stay encrypted, AI interactions are treated separately and can be analyzed for personalization without decrypting private chats. Two systems, two policies, one app. Users see one chat interface. The underlying privacy model is split.

For context on how fluid these commitments can be across a single company's product line: in the same month Meta was hardening encrypted backup infrastructure for WhatsApp, they removed optional end-to-end encryption from Instagram Direct Messages entirely, notifying users that E2EE messaging on Instagram was no longer supported. Encryption is not a permanent feature. It is an engineering and business decision that can be reversed.

How Should You Actually Evaluate a Messaging App's Privacy?

Stop asking "is it encrypted?" as a yes-or-no question. That framing collapses three separate concerns into one, and the answer will always be misleading. Instead, audit three layers:

Content encryption. Is message content end-to-end encrypted in transit? WhatsApp: yes. This is the part most people think about, and it is usually the part that works best.

Storage and backup encryption. What happens to content at rest? Is it encrypted on the device? On the cloud? By default, or only if you opt in? WhatsApp: backups are not E2EE by default. You must enable it. If you have not done so, your entire chat history may be sitting in a cloud service protected only by the cloud provider's own access controls.

Metadata and telemetry. What non-content data is collected, retained, and processed? Who you talk to, when, how often, from what device, in which groups. WhatsApp: extensive metadata collection, not covered by E2EE, used to build a social graph available to Meta.

Apply this three-part check to any product. Most apps that market themselves as "encrypted" are describing layer one only. Layers two and three are where the real exposure lives.

What Should You Do Right Now?

If you use WhatsApp and care about the privacy of your chat history, go enable encrypted backups. Settings, Chats, Chat Backup, End-to-end Encrypted Backup. Choose a password or passkey. Do it now. This is the single highest-impact action available to you, and it takes under a minute.

After that, understand what encrypted backups do not cover. Your metadata is still collected. Your social graph is still built. Your AI conversations inside WhatsApp are processed on Meta's servers (unless you use Incognito Chat where available). Online status, timestamps, group membership, call logs: all visible to WhatsApp regardless of your backup encryption setting.

If metadata exposure matters to you, the honest answer is that WhatsApp's architecture is not designed to minimize it. The Signal Protocol protects content well. The system around it collects metadata broadly. Those two things are both true at the same time.

Where Does This Leave "Encrypted = Private"?

It leaves it as an incomplete equation. Encrypted content in transit is one component of privacy. It is necessary but not sufficient. Backups, metadata, AI integrations, contact-discovery systems, and the policies governing all of them are separate surfaces, each with their own exposure profile.

The word "encrypted" on a loading screen does not tell you which of these surfaces are protected, to what degree, by default or by opt-in, and under what conditions that protection might change. The only way to know is to look at the architecture, layer by layer, and check each one against the three-part audit: content, storage, metadata.

WhatsApp gets the first one right. The second one is available but off by default. The third one is not addressed by encryption at all. That is not a condemnation of WhatsApp specifically. It is the state of almost every mainstream messaging product. The ones that claim otherwise are usually the ones you should scrutinize hardest.

If you want an AI assistant where files and transfers are zero-knowledge encrypted and the system is built around minimizing what we can see: start a free 7-day trial, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io