David Sacks criticized Anthropic for exaggerating the security threats of small models, saying that enterprise-specific small models are disrupting the business models of closed-source AI

Odaily reports: David Sacks posted on X, criticizing Anthropic's latest research report for exaggerating the security risks of small AI models autonomously replicating and spreading on IoT devices. He pointed out that the related experiments relied on devices with pre-opened debug ports, external servers, and artificially assigned propagation tasks, which is insufficient to prove that AI worms capable of autonomous spread exist in reality.

Sacks stated that traditional malware has long been able to spread at scale by exploiting vulnerabilities such as default device passwords, open ports, and lack of security updates, and that small models have not significantly changed this threat pattern.

He also mentioned that Coinbase recently fine-tuned the Qwen-3.5-9B model using its own anti-fraud data, outperforming frontier models including Anthropic Opus 4.5 in relevant tests, and reducing fraudulent transactions by approximately 30%.

Sacks believes that small models optimized for specific enterprise tasks are competitive in terms of cost, speed, and practical effectiveness, and are posing a challenge to the business models of closed-source large model companies such as Anthropic. He questioned Anthropic's emphasis on safety risks at a time when small models are beginning to gain enterprise market share, and argued that the proliferation of specialized small models may actually reduce reliance on large general-purpose AI systems.