LLMs can be forced to leak data through targeted prompt attacks, but they disclose training data only with low probability in everyday usage scenarios.
OpenAI calls for mandatory federal evaluations before AI model release but rejects regulatory approvals, positioning itself in a controlled middle ground between voluntary commitments and strict government control.
Europe’s new tech sovereignty package with four-tier certification for government IT remains vague in practical enforcement against US giants and depoliticizes the core problem of strategic dependency.
