The Trump administration is clashing with Anthropic over cybersecurity requirements that security experts view as technically unfeasible. Government officials accused the AI company of ignoring a White House cyber directive and releasing Fable 5 without proper approval. "They screwed us," one administration official told The Decoder.

The dispute centers on demands for "unhackable" large language models. The government wants guarantees that AI systems cannot be compromised or exploited, a standard that doesn't exist in current AI research or deployment. No LLM architecture today offers absolute security against all attack vectors. Vulnerabilities emerge across multiple layers: training data poisoning, prompt injection, model extraction, fine-tuning exploits, and inference-time manipulations.

Anthropic argues this requirement sets an impossible bar. The company views the demand as technically unrealistic given the current state of AI security research. Securing AI systems requires continuous iteration, adversarial testing, and architectural improvements. It's not a binary property you can guarantee upfront.

Conversations involve the Department of Commerce, the CIA, and White House science advisor Michael Kratsios. The administration appears focused on securing AI systems for national security purposes, but the framing of the requirement suggests a fundamental misunderstanding of how robustness works in machine learning.

The real issue isn't whether AI companies can make systems more secure. They can and should. But government cybersecurity frameworks often rely on absolute guarantees or certifications. LLMs don't fit that model. Security improvements happen through research, red-teaming, model updates, and deployment safeguards over time.

This tension reflects a broader gap between how government assesses risk and how AI companies operate. Traditional security audits and certification processes don't translate directly to frontier AI systems. Anthropic likely wants flexibility to deploy, iterate, and improve. Government wants assurance before deployment.

Without