Anthropic reports blocking attempts to weaponize AI models
Anthropic has disclosed that its models are being targeted for misuse in bioweapons and espionage, while new research from its Frontier Red Team reveals that AI systems are increasingly capable of performing specialized intelligence targeting and conventional weapons development tasks, posing a growing threat to national security.
According to a report from GN technics/ai (en-US), Anthropic's Frontier Red Team has introduced new evaluations demonstrating that current AI models can execute specific tactical tasks within the military 'kill chain,' such as identifying targets from fragmentary data and engineering precision drone strikes, with open-weight models also showing concerning levels of capability.
Source: GN technics/ai (en-US)According to a 154-page threat intelligence report via GN technics/ai (en-US), Anthropic disclosed that malicious actors have also leveraged its models for Russian espionage, surveillance of Uyghurs in Syria, and the development of conventional munitions like missiles and armed drones in Yemen and China. The company noted that five specific biological misuse cases involved researchers bypassing geographic safeguards to conduct work at military institutes.
Source: GN technics/ai (en-US)According to reporting from GN technics/ai (en-US), the report details a specific instance where Anthropic blocked a request to enhance the transmissibility and immune evasion of the chikungunya virus, noting that while such research has medical applications, it also poses a dual-use risk for creating more dangerous pathogens.
Source: GN technics/ai (en-US)A new report details how the company intercepted requests for cyberattacks and biological research, highlighting the growing risk of AI misuse by both individuals and state actors.
Source: GN technics/ai (en-US)
