Anthropic research shows AI worm could autonomously spread across simulated Android IoT devices
A report published by Anthropic in October, relayed by Olhar Digital (with DW), warns that AI-powered malware could install itself on IoT devices and spread autonomously. In a controlled test, researchers simulated eight Android devices. The AI model installed itself on one, replicated from it and infected the others. It reached all eight targets in four of five attempts and seven in the remaining one. Anthropic believes smaller AI models could reach this propagation capability as early as 2026, and likens the threat to a slow-growing "digital slime mold".
Detalhes técnicos
- Unlike classic worms that follow a pre-programmed script and exploit known vulnerabilities or weakly protected access points, an AI worm could solve problems on its own and adapt to different device types.
- The researchers demonstrated that AI applications can run without a large data center, on low-power devices.
- Combined with malicious software, such a worm could learn to bypass security measures, which Anthropic developers estimate would take roughly one to two days.
- Infected devices could be joined into a botnet. Each device has little processing power, but propagation speeds up as more devices are infected.
- The decentralized structure would make containment very hard, even if devices were disconnected and flaws patched.
Remediação
The article gives no specific remediation beyond the traditional measures of disconnecting devices and patching flaws, which it says would be insufficient against a decentralized AI worm.
Linha do tempo
-
Anthropic research shows AI worm could autonomously spread across simulated Android IoT devices
ler a análise desta fonte ocultar
A report published by Anthropic in October, relayed by Olhar Digital (with DW), warns that AI-powered malware could install itself on IoT devices and spread autonomously. In a controlled test, researchers simulated eight Android devices. The AI model installed itself on one, replicated from it and infected the others. It reached all eight targets in four of five attempts and seven in the remaining one. Anthropic believes smaller AI models could reach this propagation capability as early as 2026, and likens the threat to a slow-growing "digital slime mold".
Technical Details
- Unlike classic worms that follow a pre-programmed script and exploit known vulnerabilities or weakly protected access points, an AI worm could solve problems on its own and adapt to different device types.
- The researchers demonstrated that AI applications can run without a large data center, on low-power devices.
- Combined with malicious software, such a worm could learn to bypass security measures, which Anthropic developers estimate would take roughly one to two days.
- Infected devices could be joined into a botnet. Each device has little processing power, but propagation speeds up as more devices are infected.
- The decentralized structure would make containment very hard, even if devices were disconnected and flaws patched.
Remediation
The article gives no specific remediation beyond the traditional measures of disconnecting devices and patching flaws, which it says would be insufficient against a decentralized AI worm.