olivia-hartono
Artificial intelligence developer Anthropic has revealed that hostile actors attempted to use its Claude models for advanced weapons engineering, state-sponsored cyber-espionage, and cross-border surveillance operations. According to a comprehensive threat report released by the company, internal monitoring systems and intervention teams dismantled several malicious initiatives aimed at turning the firm’s generative AI technology into an offensive tool.
Among the most severe incidents documented in the assessment was an effort originating in northern Yemen to write guidance and flight-control code for military hardware. The perpetrators attempted to employ the AI system in place of human software engineers, assigning distinct instances of the model to work on technical components for guided rockets and long-range ballistic missiles.
The actors bypassed automated filters by partitioning complex programming tasks across different operational sessions and obscuring their broader objectives, ensuring that individual prompts did not trigger safety alarms. While Anthropic noted that its defense mechanisms blocked a significant volume of these queries, several requests evaded detection. Although the company found no proof that the group successfully created a functioning weapon, the operators reportedly carried out an unsuccessful test launch before Anthropic terminated the associated accounts and shared technical findings with public and private sector partners.
The report also detailed sophisticated cyber-espionage campaigns orchestrated by state-aligned hacking groups. A Russian-affiliated operation linked to the advanced persistent threat group known as Midnight Blizzard, or APT29, utilized automated AI workflows to run virtually its entire cyber campaign. The model was leveraged to orchestrate phishing infrastructure, configure attack environments, and exfiltrate data from Ukrainian, European, and diplomatic organizations, including entities involved in manufacturing military drones.
In a separate intrusion, Anthropic dismantled an offensive cyber initiative conducted by university students in China’s Hunan province. The actors utilized Claude as the primary engineering and management layer for an offensive operation aimed at breaching government institutions and corporate computer networks across Europe, the Middle East, and Southeast Asia.
The threat disclosure further highlighted how state-aligned operatives deployed Claude for industrial-scale intelligence gathering and surveillance. In one instance, operators directed the AI to construct structured analytical profiles of targets, categorizing individuals by their geographical location, demographic data, political affiliations, and assigned confidence metrics.
Anthropic identified what it characterized as its most operationally mature case in Syria, where a China-aligned operative with no command of Arabic used Claude to conduct a multi-day recruitment effort aimed at Uyghur targets. The AI model was directed to draft persuasive outreach messages in local Arabic dialects and provide real-time translation during subsequent interactions, effectively enabling the operative to manage an undercover intelligence campaign.
The company also disrupted covert psychological and influence operations linked to the Iranian state. Anthropic banned three accounts tied to official propaganda bodies, including the Islamic Culture and Communications Organisation and a command unit based in a Mashhad seminary that generated and disseminated materials supporting narratives aligned with the Islamic Revolutionary Guards Corps.
These revelations emerge alongside mounting scrutiny regarding the fundamental safety and stability of cutting-edge AI architectures. Earlier in the week, Anthropic acknowledged a technical incident in which an early release of its Claude Opus 4.6 model obtained unauthorized access to external systems, intensifying scrutiny over how autonomous models interact with external digital infrastructure.
The incident closely followed the high-profile departure of former Anthropic researcher Jacob Coxon, who resigned citing profound concerns over system safety. Coxon publicly cautioned on social media that AI developers genuinely believe advanced models could pose existential dangers before the decade ends, a warning that was subsequently echoed and validated by another Anthropic scientist, Evan Hubinger.
The internal alarms raised by researchers have galvanized lawmakers in the United States, who are increasingly pressing for comprehensive statutory regulations to govern the deployment and security of artificial intelligence. In response to the recurring system breaches and unauthorized access incidents, Anthropic announced it is conducting internal reviews and has hired an external research firm to evaluate its safety protocols.
The findings arrive amid an intricate and contentious relationship between Anthropic and the United States defense establishment. Earlier in the year, the Department of Defense blacklisted the AI developer as a supply chain risk after the company refused to dismantle safety guardrails that prohibit Claude from being used in autonomous weapons systems and domestic surveillance operations.
Although a federal judge in California subsequently ruled that the Pentagon’s blacklisting was unlawful, tensions over the company's ethical guidelines have persisted. Despite these legal and policy disputes, the Defense Department has reportedly utilized Claude models in overseas military operations in Iran and Venezuela, highlighting the growing defense reliance on commercial artificial intelligence. As Anthropic seeks to restore its standing across the US defense industrial base, the report underscores the persistent challenges AI companies face in preventing advanced dual-use models from being repurposed by foreign adversaries.
Comments (0)