Anthropic Safety Report Details AI Exploitation by State Actors, Cybercriminals, and Bioweapon Researchers

Artificial intelligence startup Anthropic has released a sweeping safety report documenting how foreign nation-states, espionage groups, and cybercriminals attempted to leverage its Claude AI model for malicious purposes over the past eight months. As reported by Mashable Tech, the threat intelligence survey categorizes security breaches across seven risk categories, detailing non-standard use cases that span biological hazards, foreign surveillance networks, cyber warfare, and conventional weapon design.

According to the report, Anthropic's security team identified five specific incidents where scientists—including individuals supported by state governments—attempted to bypass regional geographic controls to access Claude. The users disguised their research goals while prompting the AI assistant to author grant proposals involving experiments on the chikungunya virus and orthopoxviruses, the viral group that encompasses smallpox. Anthropic terminated the associated accounts and integrated the findings into its defensive systems, declining to publicly identify the researchers to prevent potential risk to their laboratories.

The safety assessment also detailed multiple cases where users attempted to harness Claude tools to build and refine physical weapons systems. In one operation attributed to a Russian freelance agent working under the project names "DronDoc" or "Serafim," the actor used Claude Code to construct software for an autonomous swarm of first-person-view kamikaze drones. In another instance, a China-based researcher focused on defense applications used the AI model to code electronic warfare modules aimed at jamming enemy communications and radar, incorporating a simulation that included 12 targets located in Taiwan.

State surveillance operations constituted another significant area of misuse, with Anthropic tracking nine distinct campaigns involving foreign entities across China, Iran, and West Africa. A operation linked to the Chinese government utilized bulk chat exports from Telegram and WhatsApp to profile and monitor Uyghur populations and journalists linked to the Syrian Army. Meanwhile, Iranian security services operated at least 16 accounts to build data-harvesting browser extensions that compromised information from 6,388 Iranian citizens, while a separate Iranian-aligned actor queried the platform to identify U.S. naval targets.

Regarding cyber warfare, the report highlighted how automated multi-agent frameworks allowed threat actors to execute reconnaissance and exploit corporate and governmental networks. A suspected Russian espionage actor designated as GTG-20006 deployed Claude to streamline phishing, domain name system hijacking, and ClickFix campaigns against Ukrainian and European government entities, foreign policy groups, and defense contractors. The actor also infiltrated three hospitality management vendors to target personal devices used by visiting Ukrainian state officials and drone manufacturers.

Anthropic noted that the broad distribution of sophisticated generative AI models has flattened capabilities across the threat landscape, providing smaller criminal syndicates with offensive tools comparable to those used by state intelligence agencies. The startup stated that publishing its threat intelligence is vital for public transparency, warning that systemic security risks will grow alongside model capabilities unless developers and defensive teams take aggressive measures to secure their platforms.

The publication of the report coincides with accelerating legislative momentum surrounding artificial intelligence governance. Anthropic recently co-signed two major AI oversight bills in California, which were backed by industry competitor OpenAI and signed into law by Governor Gavin Newsom. The legislation creates a statewide legal framework for evaluating frontier AI systems and establishes an independent registry of third-party security auditors.