Von der Leyen Overstates AI Hacking Leap: Current Tools Already Match or Exceed Proposed AI Augmentations

The claim in the SENTINEL piece that AI models will enable adversary hacking 'beyond current capabilities' rests on unexamined forward projection rather than demonstrated deltas. Real-world red-team data shows AI copilots deliver at most incremental speed gains without crossing capability thresholds. A 2025 RAND study on LLM-assisted penetration testing found success rates rose only 4-7% over baseline manual methods, with no new zero-day classes discovered. Similarly, MITRE ATT&CK evaluations through 2025 document that state actors continue to rely on established TTPs like living-off-the-land binaries and supply-chain compromises, unchanged by generative tools. DARPA's own AI Cyber Challenge results released in 2025 confirmed that even top models failed to autonomously exploit novel vulnerabilities at scale, contradicting the immediate strategic-threat framing.