WhatsApp is testing the Scam Alert security feature
Meta is launching a new WhatsApp feature in beta, called Scam Alert, which uses artificial intelligence directly on the phone to identify messages that may be part of a fraud attempt. The technology is designed so that the content of the messages does not leave the device for processing.
Online scams have become increasingly sophisticated. From messages that mimic known institutions or companies to social engineering techniques and manipulation attempts generated with the help of artificial intelligence, attackers try to convince users to disclose information, access links, or make payments.
In this context, WhatsApp is testing Scam Alert, an optional security feature based on a machine learning model that runs locally on the user’s device. According to technical documentation published by Meta on August 12, 2026, the goal is to provide a warning before the user interacts with a potential scam, while maintaining the principles of end-to-end encryption.
What is WhatsApp Scam Alert and how does it work?
Scam Alert is, at this stage, an optional feature available in a limited beta version. After activation, WhatsApp downloads a machine learning model to the phone that analyzes certain messages received from people who are not in the contact list.
The model does not simply look for certain keywords. Meta says it uses linguistic signals and conversation structure, performing probabilistic classification based on patterns observed in conversations associated with fraud. The training data comes from scam conversations previously reported by users.
If the system estimates that there is a high probability that the messages are part of a fraud attempt, the user may receive a warning in the conversation. This provides options such as blocking and reporting the sender or continuing the conversation.
It is important to note that the warning does not represent absolute proof that the message is fraudulent. It is a measure of assistance and prevention, and the user retains control over the decision.
Messages are not sent to Meta for analysis
One of the most important features of Scam Alert is that the analysis takes place directly on the phone.
According to Meta, both the machine learning model and the data from processed messages for classification remain on the device. The content of the messages is not automatically transmitted to WhatsApp, Meta, or any other third party for fraud identification.
This aspect is important because WhatsApp uses end-to-end encryption for personal conversations. A traditional fraud detection solution might involve sending messages to servers for analysis. Scam Alert seeks to avoid this approach by moving processing to the device.
Meta explicitly defines three principles for this feature:
- processing exclusively on the device;
- no automatic reporting;
- user control.
The feature can be activated or deactivated by the user, and simply detecting a possible scam does not trigger automatic transmission of the conversation to the company.
What happens if Scam Alert is wrong?
Like any machine learning-based classification system, Scam Alert can also produce false results. A legitimate message may resemble a pattern found in fraudulent conversations.
WhatsApp therefore introduces a mechanism by which the user can mark a conversation as trustworthy if the warning is deemed incorrect. In this case, the warning is removed, and Scam Alert should no longer signal that conversation in the same way.
There is also an additional option: if the user chooses to trust the conversation, they can opt to send WhatsApp the last five received messages to improve the accuracy of the system. This transmission is linked to the explicit action of the user and does not constitute automatic reporting.
How is data used without exposing encrypted conversations?
Here lies one of the interesting technical components of the project. Meta says it needs to measure whether Scam Alert is functioning correctly: how often a warning appears and what users do after it appears.
To improve Scam Alert without collecting the content of conversations, Meta uses two data protection technologies. The first (Confidential Federated Analytics) allows for distributed and secure analysis of statistics, without individual messages being sent to the company. The second (Differential Privacy) adds an additional layer of data protection, so that statistical results cannot be used to identify a specific user or their conversation.
In practice, the system only tracks aggregated indicators, such as:
- the approximate number of warnings generated;
- the categories of actions taken by users after the warning.
This data is aggregated, and the privacy mechanisms are designed so that information about an individual person cannot be deduced from the transmitted statistics. Meta states that the aggregates available to the company do not contain information about an individual user.
For processing these statistics, Meta also uses secure execution environments (Trusted Execution Environments – TEE) and confidential virtual machines. The architecture includes, among other things, verification of the executed code and mechanisms by which researchers can independently verify certain components of the system.
The system prevents directing the AI model towards specific individuals
One of the interesting risks of distributed AI systems is the possibility that a server delivers a different version of the model to a particular user.
Meta claims to have designed Scam Alert so that a specific model cannot be deliberately delivered to a predetermined user.
Model versions are recorded in a public transparency register, and each version is associated with cryptographic hashes. The register is of the append-only type, meaning that entries can be added but not modified or deleted thereafter.
Additionally, requests for downloading the model use anonymous credentials and are transmitted through an OHTTP mechanism that removes the user’s IP address from the request seen by the server delivering the model.
Another important element is that the assignment of the user to experimental variants of the model is made locally, on the device, using randomness, not by a decision made by the server for an identified person.
Users and researchers can verify which model is running
Scam Alert also introduces a verifiability component.
Meta intends to publish the necessary information for verifying the integrity of the AI models. On the device, users can enable transparency logs that show which messages have been analyzed, the result of the analysis, and the version of the model used.
In WhatsApp, this information can be accessed from Account > Request Info > Scam Alert Activity, according to the technical documentation published by Meta.
Meanwhile, security researchers are involved through the Bug Bounty program. Meta states that it provides them access to relevant components to verify whether messages remain on the device and whether the model indeed has the declared purpose of detecting scams.
When will Scam Alert be available for all WhatsApp users?
At this moment, there is no official release date yet.
Meta describes Scam Alert as a feature in a limited beta version and states that the system will be tested and improved before wide availability. The company is also working with researchers from the Bug Bounty community to identify security, privacy, and model behavior issues.
Therefore, WhatsApp users should not assume that the feature is already available on all phones or in all regions.
Why is this approach important?
Scam Alert shows an interesting direction for the security of messaging applications: using AI for protection without turning private conversations into a centralized data source for analysis.
According to BleepingComputer, WhatsApp has over 3 billion users in more than 180 countries, making any fraud prevention mechanism implemented at the platform level potentially impactful.
At the same time, users need to understand that Scam Alert does not replace personal vigilance. An AI model can identify patterns and warn about suspicious conversations, but it cannot guarantee that every fraudulent message will be detected.
The basic rule remains the same: do not send authentication codes, passwords, bank details, or other sensitive information just because a sender claims to represent a bank, a company, a friend, or an institution.
Conclusion
WhatsApp Scam Alert is an attempt to combine automated fraud detection with privacy protection. The machine learning model runs locally, messages are not automatically sent to Meta for classification, and the user decides whether a warning is correct and whether they want to report the conversation.
Moreover, Meta introduces transparency mechanisms, public registers for model versions, cryptographic verification, and testing by external researchers.
For the average user, the benefit is simple: if the feature reaches a complete release and the model proves to be accurate enough, WhatsApp could provide an additional layer of protection against scam messages, without the analysis of conversations requiring giving up end-to-end encryption.
However, for now, Scam Alert should be viewed as a system in testing, not as a guarantee against fraud. The most effective defense remains the combination of security technology, application updates, and the user’s ability to recognize the signals of a fraud attempt.
Source: engineering.fb.com
