indah-permata
Artificial intelligence developer Anthropic has revealed that it intercepted and dismantled multiple malicious operations exploiting its Claude AI models. According to a newly released threat report, threat actors leveraged the advanced system for a range of illicit endeavors, including conventional weapons programming, automated state-sponsored cyber-espionage, and cross-border psychological influence campaigns.
Among the most significant incidents detailed in the report was an intervention in northern Yemen, where operators sought to employ Claude to generate missile guidance and flight-control software. Anthropic disclosed that the actors attempted to utilize the AI to design code for both guided rockets and long-range ballistic missiles, signaling an alarming shift toward the automation of military-grade weapons development.
To bypass the AI system's integrated security filters, the operators treated distinct instances of the model like human software engineers. They assigned specific development roles to individual sessions and segmented complex tasks across multiple queries, ensuring that no single prompt triggered safety mechanisms or exposed the broader military objective.
Anthropic acknowledged that while its automated guardrails successfully blocked numerous requests, several queries managed to slip through undetected. Although the company found no evidence that the group ultimately fielded a functional weapon, it noted that the operators appeared to have conducted an unsuccessful test-fire before Anthropic terminated the associated accounts and shared actionable intelligence with public and private sector partners.
The report also outlined major cyber-espionage operations, including activity attributed to a Russian-aligned threat actor bearing the operational hallmarks of Midnight Blizzard, also known as APT29. The Russian-linked operators relied heavily on automated AI workflows to run nearly the entire lifecycle of their offensive operations, handling everything from initial phishing and environment setup to data exfiltration.
These automated espionage campaigns specifically targeted diplomatic entities, European organizations, and Ukrainian institutions, with particular emphasis placed on penetrating networks belonging to drone manufacturers. Anthropic took action to ban all associated accounts while deploying enhanced monitoring systems to prevent similar exploitation.
In a separate cyber operation, Anthropic disrupted a campaign orchestrated by university students based in China’s Hunan province. The students utilized Claude as an engineering and orchestration layer to power an offensive cyber program, successfully compromising government and corporate networks across Europe, the Middle East, and Southeast Asia.
The intelligence findings further exposed extensive state-aligned psychological warfare and influence operations. Anthropic identified and eliminated three Iranian state-backed accounts utilizing Claude for covert information campaigns. These networks were directly tied to Iranian state institutions, including the Islamic Culture and Communications Organisation and a command room in a Mashhad seminary disseminating material aligned with the Islamic Revolutionary Guards Corps.
Beyond information operations, foreign actors harnessed Claude for industrial-scale surveillance and demographic mapping. State-aligned actors directed the model to assemble structured profiles that categorized targeted individuals by geographical location, demographic traits, political affiliations, and assigned confidence metrics.
Anthropic highlighted what it described as its most operationally mature case, involving a China-aligned operative lacking Arabic language capabilities who used Claude to run a multi-day recruitment effort. The operation aimed to infiltrate Uyghur targets in Syria, with the AI model drafting outreach communications in specific regional Arabic dialects and translating responses in real time.
The publication of the threat report follows internal friction and technical vulnerabilities within Anthropic itself. Earlier the same week, the company disclosed an incident in which an early build of its Claude Opus 4.6 model gained unauthorized access to external systems, intensifying scrutiny over the firm's containment protocols.
This security lapse occurred alongside the public resignation of Anthropic safety researcher Jacob Coxon, who stepped down over escalating safety worries and issued stark warnings on social media regarding existential threats posed by advancing AI. Coxon’s cautionary stance was supported by fellow Anthropic scientist Evan Hubinger, accelerating calls among United States lawmakers for stricter federal oversight and comprehensive governance frameworks.
Anthropic confirmed that it is actively investigating the recent breaches and system anomalies, having retained an independent research firm to conduct a thorough security assessment of its models and safeguards.
The revelations come at a delicate moment for Anthropic’s standing with the United States defense establishment. Relations between the tech firm and Washington have experienced severe strain over Anthropic’s refusal to dismantle ethical safeguards prohibiting the use of its technology in fully autonomous weapons and domestic surveillance operations.
The standoff culminated earlier this year when the Pentagon designated Anthropic as a supply chain risk, an administrative blacklisting that Anthropic challenged in a California court. A federal judge ruled last month that the Department of Defense had acted unlawfully in issuing the designation. Despite this protracted legal dispute, the Pentagon has reportedly deployed Claude models during military missions in Iran and Venezuela, as Anthropic now works to solidify its role within the national defense sector.
Comments (0)