Researchers have identified critical security flaws in voice AI systems that allow attackers to embed hidden commands in audio that humans cannot hear. These adversarial attacks pose risks to voice assistants and AI-powered devices.
Voice artificial intelligence systems widely used in smartphones, smart speakers, and other devices contain exploitable vulnerabilities to inaudible audio attacks, according to recent research published in IEEE Spectrum.
The attacks work by embedding commands in ultrasonic frequencies or subtle acoustic patterns that fall below human hearing thresholds or blend into background noise. Voice AI systems, however, can detect and process these hidden instructions, potentially triggering unintended actions.
Researchers demonstrated multiple attack vectors. Some exploit the gap between what human ears perceive and what AI models process. Others use imperceptible noise patterns layered into legitimate audio. In controlled tests, attackers successfully triggered voice commands, disabled security features, and manipulated device settings without users noticing.
The vulnerability affects popular platforms including virtual assistants on phones and dedicated smart speakers. The issue stems from how voice recognition systems are trained and designed—they process a broader frequency spectrum than humans and can interpret patterns humans don't consciously register.
Security implications are significant. An attacker could theoretically inject hidden commands into videos, podcasts, or music streams to manipulate devices without user awareness. This could enable unauthorized access to accounts, trigger purchases, modify settings, or exfiltrate data.
Manufacturers and researchers are exploring defenses. Proposed solutions include improved audio filtering, anomaly detection for unusual command patterns, and enhanced user authentication for sensitive operations. Some experts advocate for hardware-level protections that limit the frequency ranges voice systems can process.
The research highlights a broader challenge in AI security: systems optimized for accuracy and sensitivity may inadvertently accept inputs humans reject as noise. As voice AI becomes more prevalent in home automation and critical infrastructure, addressing these vulnerabilities becomes increasingly important.
The findings have generated discussion in the security community about balancing voice AI usability with robust protections against adversarial audio attacks.
U.S. federal prosecutors have unsealed charges against three Russian nationals accused of operating a bulletproof hosting service that supported ransomware gangs responsible for over $62 million in damages worldwide.
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) warned that attackers are actively exploiting three vulnerabilities in Internet-exposed on-premises SharePoint Server instances. Organizations running affected versions must patch immediately.
Tailscale disclosed a critical vulnerability in its SSH implementation that allowed attackers to gain root access through insecure argument handling. The flaw has been patched in recent versions.
A new study found that social media platforms referred over 5.7 million visits to nonconsensual deepfake pornography sites between December 2025 and March 2026, with YouTube and X accounting for the majority of traffic.