Skill Squatting Attacks on Amazon Alexa
summary
Researchers send 11,460 single-word US English samples to Alexa 50 times each (573,000 transcriptions) and find 24 words yield predictable misinterpretations that are consistently mapped to the same wrong token. They build paired skills and successfully skill squatting 25 of 27 targetable errors in a developer environment, demonstrate phishing via an Amex-like skill that serves a fake login card, and show regional spear skill squatting succeeds for 8 of 11 region-specific targets.
AI-written summary · about
chapters (6)
- 0:00Introduction to Alexa skills
- 2:06The problem of skill routing
- 4:10Defining skill swatting
- 4:34Methodology and measurements
- 8:21Attack validation and impact
- 18:29Future solutions and Q&A
related talks
06mDemystifying the Vetting Process of Voice-controlled Skills on Markets
22mWhat Amazon Echo Taught Me About IoT Security
48mAlexa, Hack My Server(less) Please
05mExploring Authentication for Security-Sensitive Tasks on Smart Home Voice Assistants
48mSecurity Vulnerabilities in AI Assistant Based Applications
37mAI-Powered Slopsquatting: Is Your Software Supply Chain Compromised?
24mSecuring Language Models Against Prompt Injection
16mHarnessing AI for Malevolence: A Hackers Tale