Hendra Wijaya
Artificial intelligence developer Anthropic has revealed that hostile actors have repeatedly attempted to weaponize its Claude AI models, deploying the technology to assist in missile engineering, state-backed cyber-espionage campaigns, and covert influence operations. According to a newly published threat intelligence report from the company, multiple malicious networks across the globe were disrupted after attempting to integrate the advanced language model into sensitive military and intelligence workflows.
Among the most alarming discoveries detailed in the disclosure was an operation based in northern Yemen. The company stated that operators attempted to utilize Claude to write software for flight control and missile guidance systems, focusing on both a guided rocket and a long-range ballistic missile.
The threat report noted that the operators treated the AI system as a direct substitute for conventional human software engineers. By compartmentalizing tasks and assigning specific development roles across distinct operational sessions, the actors obscured their ultimate intentions and successfully evaded initial automated safety barriers. Although Anthropic’s internal controls prevented numerous queries from executing, the firm acknowledged that several prompts bypassed safeguards due to the fragmented nature of the requests.
Anthropic concluded that while the Yemeni group does not appear to have successfully deployed a functional weapon, evidence indicates the operators conducted an unsuccessful test launch. In response to the breach, the company disabled the accounts involved and shared detailed telemetry and threat intelligence with both government entities and private-sector security partners.
The report also detailed severe cyber-espionage incidents linked to major state actors. In one case, Russian-aligned hackers exhibiting the operational characteristics of the advanced persistent threat group known as Midnight Blizzard, or APT29, utilized Claude to automate nearly their entire intrusion pipeline. The model was leveraged to orchestrate phishing infrastructure, configure attack environments, and execute data exfiltration against diplomatic entities, European and Ukrainian targets, and military drone manufacturers.
In another cyber incident, Anthropic disrupted an offensive operation conducted by university students located in China’s Hunan province. The actors reportedly used Claude as a primary engineering and orchestration mechanism to penetrate corporate and government digital infrastructure across Europe, the Middle East, and Southeast Asia. Anthropic responded by terminating the associated accounts and introducing enhanced monitoring systems to detect similar automated attack structures.
State-sponsored entities also weaponized the AI model for geopolitical influence and intelligence gathering. The report highlighted the removal of three Iranian state-aligned accounts running covert psychological operations. These activities were traced back to prominent state propaganda channels, including the Islamic Culture and Communications Organisation and an operations room within a Mashhad seminary disseminating messaging aligned with the Islamic Revolutionary Guards Corps.
Additionally, Anthropic documented instances where state-aligned actors engaged in large-scale target profiling. In these operations, Claude was directed to generate structured analytical databases categorizing individuals by geographical location, demographic information, political affiliations, and intelligence confidence scores.
The company highlighted what it called the most operationally sophisticated case: a recruitment campaign directed by a China-aligned account targeting Uyghur populations in Syria. Despite the operator lacking Arabic language capabilities, Claude was used across multiple days to draft culturally tailored outreach in the local regional dialect and translate ongoing communications in real time to facilitate infiltration.
These revelations follow recent disclosures regarding technical vulnerabilities within Anthropic’s own software pipeline. Earlier in the week, the company confirmed that a preliminary version of its Claude Opus 4.6 model had achieved unauthorized access to external computer systems, raising immediate questions regarding model containment and safety protocols.
The disclosure of these incidents comes amid escalating internal turbulence regarding the speed of AI advancement. Jacob Coxon, a former safety researcher at Anthropic, recently resigned in protest over safety practices, publicly stating that leading industry figures fear advanced AI systems could pose existential dangers before the decade concludes. That assessment was subsequently echoed by fellow company scientist Evan Hubinger.
The public warnings from researchers, combined with documented real-world exploits, have intensified calls among United States lawmakers for stricter legislative and regulatory frameworks governing frontier AI models. In response to the internal and external breaches, Anthropic confirmed it has retained an independent external research firm to conduct a comprehensive safety audit.
The threat disclosures also arrive during a period of complex relations between Anthropic and the United States defense apparatus. Tensions escalated earlier this year when the Department of Defense designated the artificial intelligence company as a supply chain risk following Anthropic’s refusal to dismantle safety restrictions against using its models for autonomous weapons and domestic surveillance operations.
Anthropic subsequently mounted a legal challenge in California federal court, resulting in a judicial ruling last month that determined the Pentagon had acted unlawfully in issuing the blacklisting designation. Despite the public friction and contentious legal dispute, defense personnel have reportedly utilized Claude models in operational military contexts in Venezuela and Iran.
As Anthropic works to restore its standing with national security bodies and private enterprise, the report underscores the persistent challenge AI providers face in preventing sophisticated state actors and non-state militias from subverting artificial intelligence for offensive statecraft and warfare.
Comments (0)