2026-09-11 · Fuente：Tom's HW AI
Anthropic says covert accounts linked to a state sponsor used U.S. proxies and regional-block workarounds in a failed bid to get Claude to assist bioweapon research.
Anthropic says its Claude model refused instructions tied to attempts at biological weapons development, in an incident the company attributes to state-sponsored actors. According to the report, covert accounts sought to enlist the chatbot in work aimed at engineering deadlier viruses. Claude declined to comply, and Anthropic says the attempted campaign was thwarted.
The operational details underscore the sophistication of the effort. The alleged state-linked accounts routed their activity through U.S.-based proxies, a step that would mask the true origin of requests. The operators also worked to evade identification and to circumvent regional blocks — access restrictions that would otherwise have kept them out of the service. Anthropic's account indicates those countermeasures were detected and did not prevent the company from linking the activity to a suspected state sponsor.
The disclosure offers a window into how frontier AI developers monitor for misuse at scale. Anthropic identified the activity, attributed it, and says its model's safeguards held: the requests to assist with potentially developing biological weapons were refused. For an industry under pressure to demonstrate that advanced models can resist weaponization, the case functions as both a warning about adversary behavior and an argument that safety training and usage monitoring are doing their job.
For AI-compute buyers and investors, the episode highlights a dimension of the sector beyond chips and data centers: the operational cost of policing frontier deployments. State-sponsored probing, proxy-based access evasion, and attempts to extract assistance for bioweapons research are now documented parts of the threat landscape around commercial AI. The case is likely to inform ongoing debates over access controls, usage monitoring, and regional restrictions — questions that increasingly shape how frontier AI services are operated and sold.
Etiquetas: Anthropic · Claude · AI Safety · Bioweapon Research · State-Sponsored Actors · Model Misuse
萬安算交所 AIXX · AIXX · Exchange verificado · Ir al mercado