Anthropic Threat Report: How Claude Intercepted State-Sponsored Missile, Drone, and Cyber Targeting Operations

In its most exhaustive public national security disclosure to date, artificial intelligence laboratory Anthropic published a 129-page threat intelligence report detailing how foreign adversaries, state-sponsored cyber units, and militant groups attempted to exploit Claude for military targeting, missile engineering, and autonomous cyber warfare between December 2025 and August 2026.

The document introduces Anthropic's Generative Threat Group (GTG) taxonomy, chronicling dozens of adversarial campaigns disrupted across seven harm domains. Most notably, the report confirms that the commercial utility of frontier language models has progressed from general software assistance into active operational scaffolding for kinetic and digital weaponry.

Among the most alarming case studies is an operation in northern Yemen, a territorial stronghold for Iran-backed Houthi militants. According to Anthropic's investigative timeline, an engineering cell sought to use Claude Code to develop guidance, navigation, and control (GNC) software for a multistage ballistic missile and precision-guided rockets built using commercial, smartphone-class microprocessors.

The operators prompted Claude to calculate aerodynamic lift-to-drag ratios, resolve control-surface fin deflection algorithms, and write embedded C++ firmware to process sensor input from inertial measurement units. In one documented instance, the cell test-fired a guided rocket at a remote proving ground. Following an in-flight guidance failure, the militants returned to Claude within hours, pasting flight telemetry logs and sensor readouts into the model to diagnose aerodynamic instability and debug actuator commands.

While Anthropic's multi-layer safety classifiers intercepted several explicit weaponization prompts, the operators attempted to evade detection by fragmenting complex queries across multiple sessions and disguising ballistic equations as civilian academic simulations. Upon corroborating the physical testing signatures with defense analysts, Anthropic terminated all associated account clusters and shared technical forensic indicators with the US Department of Defense and international intelligence partners.

Concurrently, Anthropic disrupted a reconnaissance campaign linked to Iranian state-aligned intelligence operators. Operating amid heightened naval tensions in the Red Sea and Persian Gulf, the adversary prompted Claude to compile exhaustive tactical targeting handbooks on US Navy surface combatants, including Arleigh Burke-class guided-missile destroyers and commercial maritime vessels.

The threat actors tasked the model with analyzing open-source naval architecture blueprints, identifying hull vulnerabilities, evaluating radar cross-section blind spots, and assessing anti-torpedo defensive systems. Furthermore, the operators fed combat footage from past drone and anti-ship missile engagements into Claude's multimodal vision interface, asking the system to evaluate strike effectiveness, impact angles, and terminal guidance accuracy. Anthropic's threat intelligence team identified the anomalous query patterns and severed access before actionable fire-control solutions could be synthesized.

Beyond conventional munitions, the report details severe nation-state cyber operations executed by Russian state-sponsored actors, internally designated as GTG-20006 (aligned with Midnight Blizzard). The group integrated Claude into autonomous multi-agent attack frameworks to target over 20 defense ministries and military drone manufacturers across Ukraine and Europe, bulk-exporting mailboxes and reverse-engineering proprietary drone vision firmware.

The threat disclosure also detailed five separate incidents where credentialed scientists and bad actors attempted to use Claude for biological misuse, including protocols for synthesizing regulated pathogens and optimizing aerosolization vectors. Automated chemical and biological classifiers halted the interactions before actionable laboratory procedures were produced.

Throughout the report, Anthropic emphasizes "capability uplift"—how AI models collapse the labor and specialized knowledge gap that once separated well-funded military powers from asymmetric non-state actors. While Anthropic confirmed that none of the conventional weapons operations breached the guardrails of its newest Claude Fable or Mythos-class models, the disclosure establishes that frontier AI labs have become an active front line in global national defense.

• Detecting and countering misuse of AI: September 2026 — Anthropic Threat Intelligence
• Rebels used Anthropic's AI bot to develop guided weapons, report says — The Washington Post
• Iran used Anthropic's Claude AI to target U.S. Navy warships — Quartz
• Anthropic details bad actors' efforts to misuse its AI for bioweapons — The Guardian