Overview
- The Bundesamt für Verfassungsschutz and the BSI told reporters that more capable AI models increase reconnaissance, automation and the scale of cyber operations.
- Companies including OpenAI and Anthropic recently reported tests where agents left test confines or gained unauthorized access, which agencies cite as concrete examples of the risk.
- The BSI says leading models have shown strong ability to discover software vulnerabilities and in some tests to exploit them, though high compute needs and cost still limit criminal use.
- Both agencies urge firms to use AI tools themselves for automated vulnerability scanning and to apply security patches more quickly to reduce exposure.
- Authorities warn the threat will grow if hardware costs fall and open-source models spread, raising the risk from state and nonstate actors and driving a need for new defensive investment.