Anthropic Evaluates AI Risks in Surveillance, Weapons Development

On September 10, 2026, Anthropic’s Frontier Red Team released a detailed evaluation of how frontier AI models could be exploited for national security risks, focusing on tactical intelligence targeting and conventional weapons development. The report highlights key advances in AI capabilities that could enable surveillance, precision targeting, and autonomous weapons engineering. It also underscores the urgency of deploying safeguards to mitigate these risks.

The Red Team conducted controlled evaluations simulating military and intelligence tasks. For instance, AI models were tested on their ability to geolocate photos and social media posts, link digital identities across platforms, and even write software for drone navigation and payload delivery in simulated environments. Notably, Anthropic’s Mythos Preview and Opus 5 models demonstrated near-superhuman performance in geolocating images, with median errors as low as 37 kilometers, outperforming even elite human benchmarks in some cases.

In surveillance tasks, models showed concerning capabilities in correlating digital identities and classifying individuals. Using synthetic datasets mimicking social media activity, Mythos Preview excelled in linking multiple accounts to the same individual, even under challenging conditions. This could significantly lower the cost and expertise barriers for actors seeking to conduct mass surveillance or targeted attacks.

On the weapons front, the evaluations illustrated how AI could automate guidance and navigation tasks for drones. Opus 5, the top-performing model, successfully wrote and optimized flight control software to guide a drone through randomized, jammed GPS environments to strike a target. While the models worked within sandboxed simulations, the results emphasize the trajectory toward real-world applicability, especially as bad actors refine these capabilities.

Open-weight models, though less advanced than proprietary systems like Mythos and Opus, also demonstrated troubling capabilities. Anthropic’s report warned that such models are closing the performance gap, raising questions about the democratization of sensitive AI technologies. The findings echo prior concerns raised in Anthropic’s July 2026 pilot study on AI-controlled drones and its broader Responsible Scaling Policy evaluations.

Anthropic has implemented new safety measures, including classifiers to block misuse of its models for surveillance and weapons development. However, the report also calls for stronger regulatory and policy frameworks. It suggests that governments and developers need to address the risks posed by AI systems empowering low-resource groups with tools previously accessible only to state-level actors.

The report also raises broader geopolitical concerns. Anthropic’s CEO recently warned about the risks of authoritarian regimes exploiting AI for military and surveillance purposes, particularly if such models are trained in secrecy. The evaluations presented today provide a stark reminder of the accelerating capabilities of AI in national security contexts, with implications for both defenders and adversaries.

Looking ahead, Anthropic plans to expand its evaluations to other domains, such as space and undersea warfare. The company emphasizes the need for constant monitoring and iterative safeguards as AI models continue to advance. For policymakers and AI developers, the challenge will be balancing innovation with robust controls to prevent misuse in sensitive applications.

Image source: Shutterstock