Evaluating the Security and Privacy Risk Postures of Virtual Assistants

Borna Kalhor

and Sanchari Das

Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran

Department of Computer Science, University of Denver, Denver, Colorado, U.S.A.

Keywords:

Virtual Assistants, Privacy and Security, Vulnerability Analysis, Voice Assistants, Security Evaluation.

Abstract:

Virtual assistants (VAs) have seen increased use in recent years due to their ease of use for daily tasks. Despite

their growing prevalence, their security and privacy implications are still not well understood. To address this

gap, we conducted a study to evaluate the security and privacy postures of eight widely used voice assistants:

Alexa, Braina, Cortana, Google Assistant, Kalliope, Mycroft, Hound, and Extreme. We used three vulnerabil-

ity testing tools—AndroBugs, RiskInDroid, and MobSF—to assess the security and privacy of these VAs. Our

analysis focused on ﬁve areas: code, access control, tracking, binary analysis, and sensitive data conﬁdential-

ity. The results revealed that these VAs are vulnerable to a range of security threats, including not validating

SSL certiﬁcates, executing raw SQL queries, and using a weak mode of the AES algorithm. These vulnera-

bilities could allow malicious actors to gain unauthorized access to users’ personal information. This study

is a ﬁrst step toward understanding the risks associated with these technologies and provides a foundation for

future research to develop more secure and privacy-respecting VAs.

1 INTRODUCTION

Virtual assistants (VAs) are software systems utiliz-

ing natural language processing to facilitate human-

computer interactions via a series of intents cor-

related with service interactions and agent utter-

ances (Schmidt et al., 2018; Guzman, 2019). Cur-

rently, VAs are increasingly integrated into various

everyday interactions, such as with smartphones, the

Internet of Things (IoT), and smart speakers. Users

employ voice commands to perform a spectrum of

tasks, ranging from simple activities such as travel

planning to more intricate functions such as data anal-

ysis and information retrieval (Modhave, 2019). De-

spite the growing popularity of VAs, their adoption

hinges on the robustness of their security and privacy

features, areas that remain underexplored in existing

works. Ensuring the privacy and security of VAs is

paramount for their continued acceptance and integra-

tion into daily tasks by users.

As mentioned above, VAs come with their own se-

curity and privacy concerns since they necessitate ex-

tensive permissions to execute their designated tasks.

For example, they require access to the users’ loca-

tion for navigation, contacts for call initiation, and

storage for media playback and display (Burns and

Igou, 2019; Tan et al., 2014). Additionally, they

need calendar access for scheduling, background op-

eration permissions for continuous readiness, and net-

work access to interact with various servers (Neupane

et al., 2022). The microphone permission, crucial for

VA functionality, introduces a major concern for po-

tentially compromising user privacy (Dunbar et al.,

2021).

To delve deeper into understanding these vulner-

abilities and evaluating the security and privacy pos-

tures of VAs, we analyzed eight prevalent voice as-

sistants: Google Assistant (v0.1.474378801), Cortana

(v3.3.3.2876), Alexa (v2.2.49), Kalliope (v0.5.2),

Mycroft (v1.0.1), Braina (v3.6), Hound (v3.4.0), and

Extreme (v1.9.3). Using scanners MobSF (Abra-

ham, 2023), RiskInDroid (Georgiu, 2023), and An-

droBugs (Lin, 2023), we evaluated their security pos-

tures, scrutinized unnecessary permissions, and inves-

tigated the destinations of their network communica-

tions for potential malicious activity. This analysis

aimed to address speciﬁc research questions related

to these aspects.

• Within the contemporary technological land-

scape, what speciﬁc vulnerabilities are prevalent

in the architecture and design of leading vir-

tual assistants? Additionally, what preventative

strategies can organizations implement to mitigate

these vulnerabilities and reduce the risk of mali-

154

Kalhor, B. and Das, S.

Evaluating the Security and Privacy Risk Postures of Virtual Assistants.

DOI: 10.5220/0012389500003648

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 10th International Conference on Information Systems Security and Privacy (ICISSP 2024), pages 154-161

ISBN: 978-989-758-683-5; ISSN: 2184-4356

cious exploitation?

• As VAs become increasingly embedded in daily

digital interactions, how effectively does the ap-

plication architecture of the most widely utilized

VAs safeguard users’ sensitive data? Considering

existing literature and case studies, what practi-

cal measures or adaptations can be suggested and

veriﬁed to strengthen the intrinsic privacy and se-

curity protocols of these assistants?

Our study reveals signiﬁcant vulnerabilities in pop-

ular VAs, including improper information manage-

ment, inadequately secure or absent encryption, lack

of certiﬁcate validation, unsafe SQL query execution,

and insecure communications breaching SSL proto-

col standards. Based on the ﬁndings of this study,

we recommend that VA developers implement robust

and secure communication and access control mech-

anisms, in addition to disclosing their privacy prac-

tices. Moreover, designers and developers are en-

couraged to integrate privacy-centric design princi-

ples into VAs; this includes privacy-enhancing fea-

tures, detailed user control over data sharing, and data

collection minimization to the essentials of function-

ality (Rahman Md et al., 2022).

In the subsequent sections, we will explore the

privacy and security issues of VAs in different as-

pects, namely code, access control, binary analysis,

and tracking. The discovery of weaknesses in this

study will be conducive to growing research aimed at

fortifying the security and privacy of intelligent AI-

powered chatbots and assistants.

2 RELATED WORK

The integration of VAs into various aspects of daily

life and work has brought about signiﬁcant conve-

niences, but it has also raised substantial concerns

about privacy and security. This section reviews re-

search on VA security and privacy, connecting ﬁnd-

ings from various studies and placing them in a

broader context.

Sharif et al. provided a foundational analysis

of the privacy vulnerabilities inherent to VAs, iden-

tifying six main types of user privacy risks (Sharif

and Tenbergen, 2020). They address key concerns

such as VAs’ constant listening, which may inadver-

tently gather and store sensitive user data. They also

highlight weak authentication mechanisms in VAs,

demonstrating how malicious actors can exploit them

for unauthorized access to user data or the device.

In addition to these user-centric vulnerabilities, the

study also draws attention to the potential risks asso-

ciated with the cloud infrastructure upon which these

VAs operate, indicating that it could be susceptible to

various forms of cyberattacks (Johnston et al., 2014;

Chen et al., 2021). Complementing this, the study

conducted by Liao et al. takes a closer look at the

transparency of VA applications, particularly in terms

of how they disclose their privacy policies (Liao et al.,

2020). They discovered inconsistencies in disclosure

practices, with some applications attempting to access

sensitive data without clearly stating so in their pri-

vacy policies.

Lei et al. further expanded the security aspect of

VAs, speciﬁcally focusing on access control weak-

nesses in popular home-based VAs such as Amazon

Alexa and Google Home (Lei et al., 2018). Through

the execution of experimental attacks, they demon-

strated the feasibility of remote adversaries exploit-

ing these vulnerabilities to compromise user security.

This study underscores the critical need for stronger

authentication mechanisms to safeguard user data and

privacy. Shenava et al. extended the discussion to

smart IoT devices utilizing virtual assistant technolo-

gies, such as Amazon Alexa, to explore potential pri-

vacy violations (Shenava et al., 2022). They created

a custom Alexa skill to acquire sensitive user pay-

ment data during interactions, illustrating the serious

privacy risks associated with custom endpoint func-

tions in VA skills. This research emphasizes the ne-

cessity for robust privacy controls and stricter gover-

nance over third-party skills to protect user data on

smart IoT platforms.

Adding an empirical dimension to the discussion,

Joey et al. conducted a study to understand end-

user privacy concerns concerning smart devices (Joy

et al., 2022). Their results reveal a substantial gap

between user perceptions and the actual privacy prac-

tices of these devices, highlighting the importance of

improved privacy-preserving techniques and clearer

communication from manufacturers. Lastly, Cheng

et al. provided a literature review focused on the pri-

vacy and security challenges associated with Personal

Voice Assistants (PVAs), with a special focus on the

challenges related to the acoustic channel (Cheng and

Roedig, 2022). They introduce a taxonomy to or-

ganize existing research activities, emphasizing the

acute need for enhanced security and privacy protec-

tions in PVAs.

Weaving together the ﬁndings and insights from

these studies, we gain a comprehensive understand-

ing of the multifaceted privacy and security chal-

lenges associated with VAs. Constant listening capa-

bilities, weak authentication mechanisms, and trans-

parency issues are identiﬁed as prevalent vulnerabil-

ities and issues that need urgent attention. Further-

more, the integration of VAs with smart IoT devices

Evaluating the Security and Privacy Risk Postures of Virtual Assistants

155

introduces additional complexities and potential av-

enues for exploitation (Alghamdi and Furnell, 2023).

Addressing these challenges requires a concerted ef-

fort from researchers, developers, and manufacturers

alike, highlighting the need for stronger security pro-

tocols, clearer communication regarding data prac-

tices, and user education to foster a safer and more

secure VA ecosystem. This body of work collectively

informs the present study, providing a critical back-

drop against which we can further explore and address

the security and privacy implications of VAs.

3 METHOD

In our work, we aimed to ﬁnd security and privacy-

centric vulnerabilities prevalent in eight of the most

widely used VAs on Android, analyzing their poten-

tial exploitation of personal data for purposes such as

targeted advertising and user tracking. To select the

applications for the study, we employed a selection

method that incorporated the number of installations,

user ratings, and activity levels in open-source repos-

itories, which resulted in the inclusion of VAs such

as Google Assistant, Amazon Alexa, and Microsoft

Cortana (Warren, 2023). We used three tools for the

analysis: MobSF, RiskInDroid, and AndroBugs, each

providing a unique set of capabilities ranging from

malware analysis and API scanning to reverse engi-

neering and application sandboxing.

We speciﬁcally selected MobSF, which has been

widely used in the previous literature, due to its com-

prehensive analysis and highly active community on

GitHub. AndroBugs, armed with its efﬁcient and

robust pattern recognition, enabled us to examine

complementary vulnerable areas that were not being

examined by MobSF. RiskInDroid was chosen due

to its novel approach to using artiﬁcial intelligence

and its special focus on access control. Although

MobSF, AndroBugs, and RiskInDroid each have their

strengths, using them in conjunction provided a more

thorough vulnerability analysis.

Examining the operational mechanisms of these

tools via their GitHub repositories, we discerned that

while they all follow a generic approach of decompil-

ing APK ﬁles and juxtaposing the source code against

a predeﬁned set of rules, MobSF stands out for its

more rigorous analysis. MobSF engages in both static

and dynamic analysis of applications. In its static

analysis phase, MobSF decompiles the APK, trans-

forming it into a pseudocode that mirrors the original

source code, facilitating code examination without di-

rect access to the source. This decompiled code is

then analyzed to detect potential vulnerabilities us-

ing rules drawn from MobSF’s active open-source

community. Furthermore, MobSF scrutinizes the An-

droidManifest.xml ﬁle to extract critical information

about the application’s components and required per-

missions, shedding light on potential privacy and se-

curity concerns. The tool also evaluates the APIs em-

ployed by the application, identifying insecure APIs

as potential vulnerabilities. Additionally, MobSF

searches for hard-coded sensitive information, such as

encryption keys or credentials, which could pose sig-

niﬁcant security risks. Beyond the decompiled code,

MobSF conducts binary analysis to uncover vulner-

abilities that might only become apparent during the

compilation process (GVS, 2023). MobSF also gen-

erated a comprehensive report detailing potential vul-

nerabilities, their severity, and the speciﬁc ﬁles where

these vulnerabilities were detected.

On the contrary, RiskInDroid provides a risk in-

dex for each application, indicating the potential risk

to users, with a higher index signifying greater risk.

RiskInDroid employs classiﬁcation techniques using

machine learning libraries such as scikit − learn and

classiﬁers like Support Vector Machines (SVM) and

Multinomial Naive Bayes (MNB) to calculate a risk

score ranging from 0 to 100. RiskInDroid’s unique

capability lies in its ability to identify permissions

utilized by the application that are not explicitly de-

clared in the manifest ﬁles. AndroBugs, on the other

hand, does not generate overall assessment metrics for

the applications. Analyzing the generated reports, we

gained insights into each application’s security pos-

ture, evaluating its handling of sensitive data, utiliza-

tion of permissions, and server communications. This

enabled us to assess their resilience against potential

security threats. Our work underscores the criticality

of detecting and mitigating vulnerabilities in VAs as a

means to safeguard user privacy and security.

4 RESULTS

In our results, we discuss the vulnerabilities we found,

focusing on access control, privacy, and database se-

curity aspects, among others.

Vulnerabilities in Code. AndroBugs analysis iden-

tiﬁed several critical vulnerabilities across different

VAs. Alexa, for instance, improperly utilizes Base64

encoding as a security mechanism, leading to poten-

tial data exposure, as Base64 strings are easily de-

codable (Lei et al., 2017). Furthermore, implicit ser-

vice checking in Alexa, Cortana, and Hound could

allow unauthorized access and execution of sensi-

tive functions, undermining system security. Alexa

ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy

156

Table 1: Overview of Critical Security Vulnerabilities in Virtual Assistants.

Critical Vulnerability Virtual Assistants

Non-SSL URLs Amazon, Cortana, Extreme, Hound,

Kalliope, Mycroft

Implicit Intent Usage Alexa, Cortana, Extreme, Hound, My-

croft

Missing Stack Canary in Shared Objects Alexa, Cortana, Hound, Mycroft

Permissive HOSTNAME VERIFIER Cortana, Extreme, Kalliope

SSL Certiﬁcate Validation Bypass Cortana, Extreme, Kalliope

Raw SQL Queries in SQLite Cortana, Kalliope

Runtime Command Execution Cortana, Extreme

WebView Javascript Interface Vulnerability (CVE-2013-4710) Cortana, Extreme

ContentProvider Exported Alexa

Base64 Encoding Used as Encryption Alexa

Misconﬁgured ”intent-ﬁlter” Cortana

ECB Mode AES Usage Cortana

Clear Text Trafﬁc Enabled Cortana

Non-Standard Activity Launch Mode Cortana

CBC Mode AES with Padding Oracle Extreme

Standhogg 2.0 Vulnerability Extreme

System-Level Permissions in Manifest Extreme

Insecure SSL Pinning with Byte Array/Hard-Coded Cert Info Kalliope

Fragment Vulnerability (CVE-2013-6271) Mycroft

Debug Mode Enabled Mycroft

Dangerous Sandbox Permissions Mycroft

also exposes its ContentProvider, risking unautho-

rized data access, and communicates over non-SSL

URLs, exposing data to potential DNS hijacking at-

tacks (Shahriar and Haddad, 2014). Extreme’s sus-

ceptibility to Strandhogg 2.0 allows malicious apps

to impersonate legitimate ones, jeopardizing user se-

curity. To mitigate this, setting ’singleTask’ or ’sin-

gleInstance’ launch modes in AndroidManifest.xml is

recommended, ensuring single activity instances and

thwarting malicious activity duplication.

In our analysis of Cortana and Kalliope, we iden-

tiﬁed several SSL security vulnerabilities. Both ap-

plications permit a user-deﬁned HOSTNAME VERI-

FIER to indiscriminately accept all Common Names

(CN), constituting a critical vulnerability that en-

ables Man in the Middle attacks using a valid certiﬁ-

cate (MITRE, 2023). Furthermore, Cortana initiates

connections with 13 URLs that are not secured by

SSL. The application also fails to validate SSL certiﬁ-

cates properly, accepting self-signed, expired, or mis-

matched CN certiﬁcates. Additionally, the ”intent-

ﬁlter” in the app is misconﬁgured, lacking associated

”actions,” thereby rendering the component inopera-

tive (Developers, 2023).

Our investigation also uncovered a potential for

Remote Code Execution due to a critical vulnerabil-

ity in the WebView ”addJavascriptInterface” feature,

allowing JavaScript to control the host application.

This is particularly hazardous for applications target-

ing API LEVEL JELLYBEAN (4.2) or lower, as it ex-

poses the application to arbitrary Java code execution

with the host application’s permissions (Labs, 2013).

In Mycroft, we discovered numerous critical vulnera-

bilities, including implicit service checking and three

URLs that are not secured by SSL. The application

is susceptible to the Fragment Vulnerability (CVE-

2013-6271), enabling attackers to execute arbitrary

code in the application context (NIST, 2013). Ad-

ditionally, the ”debuggable” property is set to ”true”

in AndroidManifest.xml, exposing debug messages to

attackers via Logcat.

Access Control and Application Permissions.

Commonly, VAs require access to sensitive informa-

tion such as the device’s location, storage, and inter-

net connectivity (Alepis and Patsakis, 2017). How-

ever, it poses a challenge for researchers to ascer-

tain the exact utilization of speciﬁc permissions by

developers. This ambiguity renders VAs attractive

targets for malicious actors, who could exploit these

permissions under the guise of legitimate functional-

ity (Bolton et al., 2021). Identifying an application’s

access permissions is relatively straightforward. One

can inspect the AndroidManifest.xml ﬁle to review

the declared permissions necessary for the app’s func-

tionality. Additionally, automated open-source code

analysis tools, such as MobSF and RiskInDroid, can

reverse engineer and scrutinize the code to identify

various permissions. Employing the Android SDK

also enables the monitoring of application requests to

determine the assets being accessed.

Our study revealed that Extreme uti-

lizes a system-level, special permission,

”W RIT E SECURE SET TINGS”, as declared

in its Android Manifest ﬁle. This signature-level per-

mission is typically reserved for applications signed

with the system’s signing key, and it grants the app

extensive access, allowing it to modify the system’s

Evaluating the Security and Privacy Risk Postures of Virtual Assistants

157

Table 2: Trackers utilized by VAs for various purposes. No trackers were found in the remaining VAs.

Tracker Name ⧹ VA name Alexa Cortana Extreme Hound

Amazon Analytics (Amazon in-

sights)

Analytics

Bugsnag Crash reporting

Google Firebase Analytics Analytics Analytics Analytics Analytics

Metrics Analytics

Adjust Analytics

Google AdMob Advertisement

Google CrashLytics Crash reporting Crash reporting

Huawei Mobile Services (HMS)

Core

Analytics, Ad-

vertisement, Lo-

cation

Facebook Analytics Analytics

Facebook Login Identiﬁcation

Facebook Share [Unspeciﬁed]

Google Analytics Analytics

Houndify Identiﬁcation,

Location

Localytics Analytics, Pro-

ﬁling

OpenTelemetry (OpenCensus,

OpenTracing)

Analytics

secure settings. Such capabilities present substantial

security risks. Generally, special permissions of

this nature are granted exclusively to preloaded

applications, those integral to the device’s operating

system, or apps installed by Google. Furthermore,

the ”REQUEST INSTALL PACKAGES” permission

observed in Extreme permits the application to

request the installation of additional packages. This

functionality raises security concerns, as seemingly

benign applications might exploit this permission to

deceive users into installing malicious packages.

Tracking in VAs. Trackers are software compo-

nents integrated into applications to collect data on

users’ activities, capturing details such as the duration

of app usage, clicked ads, and more. Often provided

by third parties, these trackers can be used to predict

future user behavior and interests. Companies utilize

this information to tailor their products and services to

user needs. Still, trackers can be seen as invasive, rais-

ing privacy concerns among users and privacy advo-

cates (Kollnig et al., 2021). Nevertheless, developers

sometimes employ trackers for benevolent purposes,

such as application maintenance. Trackers assist in

identifying the most and least effective parts of an

application, guiding necessary changes, and are also

commonly used for crash reporting (Farzana et al.,

2018).

Most applications opt not to develop their own

tracking systems, instead utilizing established track-

ers provided by major companies. Our study indicates

that a signiﬁcant number of test subjects employ this

technology for a variety of purposes. Google Firebase

Analytics emerged as the predominant tracker among

the applications studied, primarily utilized for analyt-

ics. Google CrashLytics is another prevalent tool used

for transmitting crash reports to developers. Addition-

ally, various tracking products developed by Amazon,

Google, Facebook, Twitter, Huawei, and others are

employed for multiple purposes, including analytics,

crash reporting, and user behavior prediction. Table 2

outlines the trackers found in the VAs analyzed in this

study and their speciﬁc purposes.

Binary Analysis of Shared Libraries. Each appli-

cation undergoes numerous repetitive processes and

operations beyond its core functionality. Developers

often maximize code reusability across different ap-

plications by relying on pre-developed code compo-

nents, such as functions, classes, and variables, col-

lectively referred to as shared libraries. Although

this practice offers signiﬁcant advantages in terms of

reusability, it also introduces potential security risks.

Malicious actors can exploit vulnerabilities in out-

dated or improperly maintained libraries to compro-

mise the system. A notable example occurred in

November 2021 when security researchers discovered

a critical vulnerability in a widely-used Java logging

framework, a vulnerability that had remained unde-

tected since 2013—known as the log4shell vulnera-

bility (CVE-2021-44228) (Wortley et al., 2021).

Sensitive Data Conﬁdentiality. In our analysis, we

discovered vulnerabilities in several VAs. Braina and

Google Assistant are susceptible to CWE-532 and

CWE-276, meaning they may inadvertently leak sen-

sitive information into log ﬁles or expose such data

through other means. Braina additionally poses a se-

curity risk due to its read/write operations on external

storage, which could be accessed by other applica-

tions. Cortana employs the ECB mode for crypto-

graphic encryption, a method considered weak since

ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy

158

Figure 1: App-Based Security Scores Generated by MobSF.

identical plaintexts yield identical ciphertexts (CWE-

327), resulting in predictable ciphertexts. Further-

more, Cortana utilizes the MD5 hashing algorithm,

which is known for its susceptibility to collisions. Our

ﬁndings also indicate that Alexa may expose secrets,

such as API keys. Kalliope, on the other hand, is at

risk of SQL injection due to its unsafe execution of

raw SQL queries, potentially enabling the execution

of malicious queries by users. Hound’s domain con-

ﬁguration allows clear text trafﬁc to speciﬁc domains,

creating a security vulnerability. Google Assistant

uses an insecure random number generator (CWE-

330) and employs the MD5 algorithm for hashing.

Extreme is signed with a v1 signature scheme, render-

ing it vulnerable to the Janus vulnerability on Android

versions 5.0-8.0.

The comprehensive security scores of the apps, as

generated by MobSF, are depicted in Figure 1. We

should note that these scores provide insights into the

number and severity of weaknesses identiﬁed by the

tool only.

5 DISCUSSION

The pervasive integration of VAs into daily life un-

derscores the crucial need for enhanced security and

privacy measures. Our study elucidates a spectrum

of vulnerabilities in widely-used VAs, predominantly

stemming from insecure communications, misconﬁg-

urations, and suboptimal encryption practices. These

vulnerabilities not only expose users to potential ex-

ploitation but also raise substantial privacy concerns,

particularly given the observed instances of user ac-

tivity tracking and data transmission to third-party

services.

This study reveals vulnerabilities in several preva-

lent VAs, could assist corporations in pinpointing de-

ﬁcient aspects of their software and striving to rein-

force them. Additionally, it educates the public on the

risks faced by the current status of VAs regarding han-

dling their data and how companies use their data. It

also serves as a reminder that the esteemed renown of

multinational corporations does not necessarily guar-

antee the security of their products. Based on this

study, we advocate for the adoption of privacy-by-

design principles (Sokolova et al., 2014) and compre-

hensive security practices across the software devel-

opment life cycle (SDLC)(Assal and Chiasson, 2018).

Speciﬁcally, organizations should implement end-to-

end encryption for sensitive data, both in transit (us-

ing TLS) and at rest (employing robust encryption

schemes such as AES-256), and refrain from embed-

ding sensitive information like API keys directly in

the code. Regular patching for software and depen-

dencies is paramount, as many exploits target known

vulnerabilities for which ﬁxes are readily available.

For developers, embracing secure communication

protocols such as HTTPS is imperative to ensure the

integrity and conﬁdentiality of data exchanges be-

tween VAs and servers (Hadan et al., 2019). The inte-

gration of robust authentication mechanisms, includ-

ing multi-factor and biometric authentication, adds a

critical layer of security, safeguarding user data from

unauthorized access (Khattar et al., 2019; Das et al.,

2018; Das et al., 2020). Transparent and user-friendly

privacy policies are vital, enabling users to make in-

formed decisions regarding VA usage. Regular up-

dates to these policies ensure alignment with cur-

rent data collection and usage practices (Xie et al.,

2022). Users, in turn, must be vigilant and proac-

tive in safeguarding their privacy. This includes scru-

tinizing the permissions requested by VAs, employ-

ing strong, unique passwords, and considering multi-

factor authentication (Warkentin et al., 2011; Jensen

et al., 2021). Moreover, explicit user consent should

be a prerequisite for the collection of sensitive data.

VAs should empower users with greater control over

their data, including the ability to delete it, dictate its

retention duration, or opt out of data collection en-

tirely (Hutt et al., 2023). Addressing these challenges

necessitates a concerted effort from developers, users,

and regulatory bodies aimed at establishing and main-

taining a secure and privacy-respecting virtual envi-

ronment.

6 LIMITATIONS & FUTURE

WORK

Due to proprietary restrictions imposed by Apple, iOS

assistants were not within our investigatory purview.

Additionally, there is a limited array of widely used

VAs available, generally pre-installed on devices,

which inherently narrows the scope of this work.

While this study successfully identiﬁes vulnerabili-

ties, we plan to conduct subsequent evaluations to

augment these ﬁndings through experimental vali-

Evaluating the Security and Privacy Risk Postures of Virtual Assistants

159

dation. Furthermore, this study predominantly em-

ployed static analysis. In our future work, we plan to

incorporate dynamic and hardware analysis, utilizing

single-board computers to deploy the assistants.

7 CONCLUSION

The proliferation of VAs entails substantial privacy

and security ramiﬁcations for end-users. This study

assessed the security postures and privacy implica-

tions of the eight most commonly used VAs, en-

compassing Alexa, Braina, Cortana, Google Assis-

tant, Kalliope, Mycroft, Hound, and Extreme. Utiliz-

ing automated vulnerability scanners, we conducted

static code analysis to uncover latent code vulnera-

bilities and assess the necessity of requested permis-

sions. Our ﬁndings reveal that numerous VAs are rid-

dled with critical vulnerabilities, potentially exposing

users’ private data to malicious actors. Many VAs ex-

hibit ﬂawed information handling practices and em-

ploy suboptimal encryption standards. Our analy-

sis also showed the prevalence of third-party track-

ers within these applications, highlighting potential

data-sharing practices with advertising entities. We

underscore the imperative for stringent security mea-

sures, advocating for enhanced encryption practices

and robust authentication mechanisms. Additionally,

we encourage users and developers alike to cognize

the inherent privacy risks associated with VA usage,

fostering a security-conscious virtual environment.

REFERENCES

Abraham, A. (2023). Mobsf/mobile-security-framework-

mobsf. https://github.com/MobSF/Mobile-Security-F

ramework-MobSF.

Alepis, E. and Patsakis, C. (2017). Monkey says, monkey

does: security and privacy on voice assistants. IEEE

Access, 5:17841–17851.

Alghamdi, S. and Furnell, S. (2023). Assessing security

and privacy insights for smart home users. In ICISSP,

pages 592–599.

Assal, H. and Chiasson, S. (2018). Security in the soft-

ware development lifecycle. In Fourteenth symposium

on usable privacy and security (SOUPS 2018), pages

281–296.

Bolton, T., Dargahi, T., Belguith, S., Al-Rakhami, M. S.,

and Sodhro, A. H. (2021). On the security and privacy

challenges of virtual assistants. Sensors, 21(7):2312.

Burns, M. B. and Igou, A. (2019). “alexa, write an audit

opinion”: Adopting intelligent virtual assistants in ac-

counting workplaces. Journal of Emerging Technolo-

gies in Accounting, 16(1):81–92.

Chen, C., Mrini, K., Charles, K., Lifset, E., Hogarth, M.,

Moore, A., Weibel, N., and Farcas, E. (2021). To-

ward a uniﬁed metadata schema for ecological mo-

mentary assessment with voice-ﬁrst virtual assistants.

In Proceedings of the 3rd Conference on Conversa-

tional User Interfaces, pages 1–6.

Cheng, P. and Roedig, U. (2022). Personal voice assistant

security and privacy—a survey. Proceedings of the

IEEE, 110(4):476–507.

Das, S., Dingman, A., and Camp, L. J. (2018). Why johnny

doesn’t use two factor a two-phase usability study of

the ﬁdo u2f security key. In Financial Cryptography

and Data Security: 22nd International Conference,

FC 2018, Nieuwpoort, Curac¸ao, February 26–March

2, 2018, Revised Selected Papers 22, pages 160–179.

Springer.

Das, S., Kim, A., Jelen, B., Huber, L., and Camp, L. J.

(2020). Non-inclusive online security: older adults’

experience with two-factor authentication. In Pro-

ceedings of the 54th Hawaii International Conference

on System Sciences.

Developers, A. (2023). Intents and intent ﬁlters — Android

Developers — developer.android.com. https://develo

per.android.com/guide/components/intents-filters.

[Accessed 26-10-2023].

Dunbar, J. C., Bascom, E., Boone, A., and Hiniker, A.

(2021). Is someone listening? audio-related pri-

vacy perceptions and design recommendations from

guardians, pragmatists, and cynics. Proceedings of

the ACM on Interactive, Mobile, Wearable and Ubiq-

uitous Technologies, 5(3):1–23.

Farzana, B., Muhammad, D., Salman, A., Zia, U., and

Haider, A. (2018). Accident detection and smart res-

cue system using android smartphone with real-time

location tracking. International Journal of Advanced

Computer Science and Applications 9 no.

Georgiu, G. C. (2023). Claudiugeorgiu/riskindroid. https:

//github.com/ClaudiuGeorgiu/RiskInDroid.

Guzman, A. L. (2019). Voices in and of the machine:

Source orientation toward mobile virtual assistants.

Computers in Human Behavior, 90:343–350.

GVS, C. (2023). Binary code analysis vs source code anal-

ysis. https://www.appknox.com/blog/binary-code-a

nalysis-vs-source-code-analysis.

Hadan, H., Serrano, N., Das, S., and Camp, L. J. (2019).

Making iot worthy of human trust. In TPRC47: The

47th Research Conference on Communication, Infor-

mation and Internet Policy.

Hutt, S., Das, S., and Baker, R. S. (2023). The right to

be forgotten and educational data mining: Challenges

and paths forward. International Educational Data

Mining Society.

Jensen, K., Tazi, F., and Das, S. (2021). Multi-factor

authentication application assessment: Risk assess-

ment of expert-recommended mfa mobile applica-

tions. Proceeding of the Who Are You.

Johnston, M., Chen, J., Ehlen, P., Jung, H., Lieske, J.,

Reddy, A., Selfridge, E., Stoyanchev, S., Vasilieff, B.,

and Wilpon, J. (2014). Mva: The multimodal virtual

assistant. In Proceedings of the 15th Annual Meeting

ICISSP 2024 - 10th International Conference on Information Systems Security and Privacy

160

of the Special Interest Group on Discourse and Dia-

logue (SIGDIAL), pages 257–259.

Joy, D., Kotevska, O., and Al-Masri, E. (2022). Investigat-

ing users’ privacy concerns of internet of things (iot)

smart devices. In 2022 IEEE 4th Eurasia Conference

on IOT, Communication and Engineering (ECICE),

pages 70–76. IEEE.

Khattar, S., Sachdeva, A., Kumar, R., and Gupta, R. (2019).

Smart home with virtual assistant using raspberry pi.

In 2019 9th International Conference on Cloud Com-

puting, Data Science & Engineering (Conﬂuence),

pages 576–579. IEEE.

Kollnig, K., Dewitte, P., Van Kleek, M., Wang, G., Omeiza,

D., Webb, H., and Shadbolt, N. (2021). A fait accom-

pli? an empirical study into the absence of consent

to {Third-Party} tracking in android apps. In Sev-

enteenth Symposium on Usable Privacy and Security

(SOUPS 2021), pages 181–196.

Labs, W. (2013). WebView addJavascriptInterface Remote

Code Execution — labs.withsecure.com. https://labs

.withsecure.com/publications/webview-addjavascript

interface-remote-code-execution. [Accessed 26-10-

2023].

Lei, L., He, Y., Sun, K., Jing, J., Wang, Y., Li, Q., and

Weng, J. (2017). Vulnerable implicit service: A re-

visit. In Proceedings of the 2017 ACM SIGSAC Con-

ference on Computer and Communications Security,

pages 1051–1063.

Lei, X., Tu, G.-H., Liu, A. X., Li, C.-Y., and Xie, T. (2018).

The insecurity of home digital voice assistants-

vulnerabilities, attacks and countermeasures. In 2018

IEEE conference on communications and network se-

curity (CNS), pages 1–9. IEEE.

Liao, S., Wilson, C., Cheng, L., Hu, H., and Deng, H.

(2020). Measuring the effectiveness of privacy poli-

cies for voice assistant applications. In Annual Com-

puter Security Applications Conference, pages 856–

869.

Lin, Y.-C. (2023). Androbugs/androbugs framework. https:

//github.com/AndroBugs.

MITRE (2023). CWE - CWE-297: Improper Valida-

tion of Certiﬁcate with Host Mismatch (4.13) —

cwe.mitre.org. https://cwe.mitre.org/data/definitio

ns/297.html. [Accessed 26-10-2023].

Modhave, S. (2019). A survey on virtual personal assistant.

International Journal for Research in Applied Science

and Engineering Technology, 7(12):305–309.

Neupane, S., Tazi, F., Paudel, U., Baez, F. V., Adamjee,

M., De Carli, L., Das, S., and Ray, I. (2022). On the

data privacy, security, and risk postures of iot mobile

companion apps. In IFIP Annual Conference on Data

and Applications Security and Privacy, pages 162–

182. Springer.

NIST (2013). Nvd - cve-2013-6271. https://nvd.nist.gov/v

uln/detail/CVE-2013-6271. Accessed on October 27,

2023.

Rahman Md, R., Miller, E., Hossain, M., and Ali-Gombe,

A. (2022). Intent-aware permission architecture: A

model for rethinking informed consent for android

apps [intent-aware permission architecture: A model

for rethinking informed consent for android apps].

ICISSP 2022.

Schmidt, B., Borrison, R., Cohen, A., Dix, M., G

artler,

M., Hollender, M., Kl

opper, B., Maczey, S., and Sid-

dharthan, S. (2018). Industrial virtual assistants: Chal-

lenges and opportunities. In Proceedings of the 2018

ACM International Joint Conference and 2018 In-

ternational Symposium on Pervasive and Ubiquitous

Computing and Wearable Computers, pages 794–801.

Shahriar, H. and Haddad, H. M. (2014). Content provider

leakage vulnerability detection in android applica-

tions. In Proceedings of the 7th International Confer-

ence on Security of Information and Networks, pages

359–366.

Sharif, K. and Tenbergen, B. (2020). Smart home voice

assistants: a literature survey of user privacy and se-

curity vulnerabilities. Complex Systems Informatics

and Modeling Quarterly, 1(24):15–30.

Shenava, A. A., Mahmud, S., Kim, J.-H., and Sharma,

G. (2022). Exploiting security and privacy issues

in human-iot interaction through the virtual assistant

technology in amazon alexa. In International Con-

ference on Intelligent Human Computer Interaction,

pages 386–395. Springer.

Sokolova, K., Lemercier, M., and Boisseau, J.-B. (2014).

Privacy by design permission system for mobile ap-

plications. In The Sixth International Conferences on

Pervasive Patterns and Applications, PATTERNS.

Tan, J., Drolia, U., Gandhi, R., and Narasimhan, P. (2014).

Poster: Towards secure execution of untrusted code

for mobile edge-clouds. ACM WiSec.

Warkentin, M., Johnston, A. C., and Shropshire, J. (2011).

The inﬂuence of the informal social learning environ-

ment on information privacy policy compliance efﬁ-

cacy and intention. European Journal of Information

Systems, 20(3):267–284.

Warren, T. (2023). Using cortana on ios or android. https:

//www.theverge.com/2021/4/1/22361687/microsoft-c

ortana-shut-down-ios-android-mobile-app. Cortana’s

support on mobile ended on March 31, 2021.

Wortley, F., Allison, F., and Thompson, C. (2021).

Log4shell: Rce 0-day exploit found in log4j, a pop-

ular java logging package — lunatrace.

Xie, F., Zhang, Y., Yan, C., Li, S., Bu, L., Chen, K., Huang,

Z., and Bai, G. (2022). Scrutinizing privacy policy

compliance of virtual personal assistant apps. In 37th

IEEE/ACM International Conference on Automated

Software Engineering, pages 1–13.

Evaluating the Security and Privacy Risk Postures of Virtual Assistants

161