Artificial Intelligence System Used to Develop Advanced Missile Software
Sanaa, Saturday, 12 September 2026.
In an unprecedented safety breach, a weapons cell substituted artificial intelligence for human software engineers, utilizing the platform to design and troubleshoot guidance systems for multiple advanced rocket programs.
Anthropic’s Investigation Findings
On 2026-09-10, Anthropic published a threat intelligence report detailing the disruption of these activities, noting that the cell assigned different roles to multiple, simultaneous Claude instances to mimic a development team [1][3]. The actors successfully concealed their objectives by fragmenting the project across multiple conversations, though Anthropic eventually detected the activity through an internal investigation and banned the involved accounts [1]. While Anthropic found no evidence of an operational weapon produced via Claude, the cell conducted a real-world test-fire of a guided rocket in Yemen that failed [1][3]. The actors subsequently used Claude to troubleshoot the failure, attempting to resolve the error within hours of the incident [1][3]. This specific cell represents one of six total conventional-weapons investigations identified by Anthropic involving actors in Yemen, China, and Russia in September 2026 [1]. Mathematically, this single Yemen-based operation constitutes 16.667 percent of the total conventional weapons investigations disclosed in the report [1]. The cell developed an offline simulation toolkit to reduce dependence on Claude and external software like MATLAB, indicating a strategy to bypass reliance on continuous AI access [1][3].
Technical Capabilities and Weapon Programs
The weapons-development cell utilized Claude Code to engineer guidance software for rockets and missiles, effectively substituting the AI platform for human software engineers [1][2]. The group focused on three specific weapons programs: a guided rocket, a multi-stage ballistic missile with a range exceeding 2,000 km, and a proposed hypersonic glide vehicle [1][3]. Anthropic reported that new evaluations indicate frontier AI models are demonstrating consistent progress in performing military and intelligence tasks previously requiring human expertise [1]. The guided rocket utilized a commodity phone-class flight computer with final-phase homing guidance, highlighting the accessibility of dual-use technology [3]. Despite existing safety guardrails meant to block illicit requests, the group effectively bypassed these measures through fragmented prompting [1][2].
Security Implications and Response
Anthropic has since shared intelligence regarding this operation with unnamed public and private-sector partners to mitigate future risks [1][4]. The revelation highlights growing national security concerns and sovereign regulatory risks for AI developers as advanced dual-use models become increasingly accessible to sanctioned non-state actors globally [1][2]. The company stated it covers how people tried to misuse Claude for cyberattacks, influence operations, surveillance, biology, and building weapons, and how they found and stopped them [1]. Anthropic disrupted every operation in the report, though the accessibility of such models remains a persistent challenge for the industry [1][4]. The incident underscores the need for robust monitoring systems capable of detecting fragmented misuse patterns across multiple user sessions [2][3].