Anthropic's Frontier Red Team has developed new evaluations to measure AI capabilities in tactical intelligence targeting and conventional weapons development. These evaluations assess how AI progress is changing the risk landscape across different parts of the "kill chain," which are end-to-end conceptual models of engagements like "find, fix, track, target, engage, assess."
A new report from Anthropic's Threat Intelligence Team indicates that AI models are already perceived as beneficial by threat actors for misuse in surveillance and conventional weapons development. The Frontier Red Team's complementary capability evaluations show that models are making consistent progress on simulated intelligence and weapons development tasks.
Key Points
- Anthropic's Frontier Red Team developed evaluations for tactical intelligence targeting and conventional weapons development.
- Tactical intelligence targeting includes identifying and fixing targets of interest, a process historically labor-intensive.
- Conventional weapons development involves tasks like engineering drones to strike a moving target.
- The evaluations show models making consistent progress on simulated intelligence and weapons development tasks.
- Open-weights models tested are behind the frontier (typically between Sonnet and Mythos-class models in performance) but can still achieve concerning capability levels.
- The Mythos Preview model is the top performer in account linkage and individual classification tasks in simulated social media content.
- Kimi K3 performs similarly to frontier models on easier tasks but lags on more difficult samples for account linkage.
Context
According to Anthropic Research, cybersecurity and biorisk are well-studied domains of AI misuse risk, but modern conflict often occurs in more conventional realms. The report suggests that AI's progress in data analysis, software development, and coding can apply to specialized national security domains. Historically, improvements in the intelligence targeting process required expert human labor and judgment, but AI could make this labor less scarce and more widely available.
Why It Matters
These findings indicate that AI models, even those short of the frontier, have potential intelligence and military applications, which could affect how models are trained, safeguarded, and released. Builders and deployers should note the potential for AI to contribute to breakthroughs in these domains, influencing future development and release strategies.
What To Do
- Review the full Anthropic Research report to understand the methodologies and detailed results of the evaluations.
- Note the performance differences between frontier models like Mythos Preview and open-weights models in intelligence targeting tasks.
- Consider the implications of AI's progress in making intelligence targeting labor less scarce for potential misuse scenarios.
- Watch for further research or guidance from Anthropic regarding the training, safeguarding, and release of models with these capabilities.