Experts Warn of 'No Magic Kill Switch' as US Adversaries Exploit AI to Build Weapons

Artificial intelligence safety experts have warned that large AI models lose control of sensitive data once adversaries extract it to develop weapon systems, highlighting the absence of a 'magic kill switch.' According to David Reis, head of geopolitical risk at Alice, analysis of AI abuse shows that even when information is taken offline, it remains vulnerable to exploitation by US adversaries. He pointed to Yemen's Houthis using AI platforms for rocket development as proof of accelerating weapons capabilities and the limitations of tech company safeguards.

A recent report by Anthropic revealed that a Yemen-based weapons cell utilized Claude to build guidance, navigation, and control software for a precision rocket, alongside work on ballistic missile and hypersonic-glide systems. The Houthis bypassed safety features by concealing their objectives and splitting tasks across multiple chat sessions. Anthropic noted that the rebel group eventually test-fired a guided rocket, though an operational weapon has not yet been successfully fielded.

Mr Reis highlighted the sophistication of the offline simulation toolkit developed by the Houthis. His employer, Angel, offers a security product named Rabbit Hole that flags malicious interactions. However, once the Houthis took their custom system offline, investigators lost track of it. While the industry has praised Anthropic for publishing these details, Mr Reis emphasized that no simple technological kill switch exists to neutralize AI capabilities once transferred offline. Former Meta executive and ex-deputy prime minister of the UK, Nick Clegg, echoed these concerns, stating he knew of no plausible way to regain control over decentralized data centers and global servers.

Mr Reis explained that the low cost of experimenting with frontier AI gives cyber attackers little incentive to stop trying to circumvent safeguards. The Houthis used obfuscation tactics—such as hiding goals and splitting workflows—that are relevant to various threat actors engaging in cyber operations, influence campaigns, and surveillance. He introduced the concept of “uplift,” noting how AI dramatically improves the scale, depth, and autonomy of these operations.

Furthermore, the Anthropic report detailed how Iranian-linked operations utilized Claude for intelligence and influence activities to plan campaigns and draft operational manuals. Mr Reis stated that intelligence agencies like Iran's would naturally exploit the technology, though this could also create new tracking opportunities for western security services to monitor groups like the Houthis and Iran's Islamic Revolutionary Guard Corps. Ultimately, experts agree that security must rely on layered defenses and continuous monitoring rather than a single technological safeguard.