About half a dozen Anthropic engineers are embedded with the NSA, helping deploy the company's AI for cyber operations under a program called Mythos. This is the most direct evidence yet of frontier AI companies working inside the intelligence community.
These aren't consultants visiting occasionally. They're forward-deployed engineers working directly with NSA personnel on operational systems. And it raises questions that should make everyone uncomfortable: what guardrails exist when the most powerful AI systems are deployed for offensive cyber operations?
Anthropic publicly positions itself as the safety-focused AI company. They write papers about constitutional AI, publish research on interpretability, and just called for an industry-wide development pause. But they're also embedding engineers in the intelligence apparatus.
There's nothing inherently wrong with this. The NSA has legitimate defensive responsibilities. Cyber threats are real and sophisticated. AI could meaningfully improve threat detection and response. But "Mythos cyber operations" sounds less like defense and more like offense.
From my experience in tech, forward deployment means these engineers are deeply integrated. They're not just providing API access - they're customizing systems, troubleshooting edge cases, potentially building capabilities specific to NSA requirements.
What happens when those requirements conflict with Anthropic's stated principles? If the NSA wants Claude to help develop exploits or conduct offensive operations, do the engineers push back? Do they have authority to refuse? Or are they there to make the technology work regardless of use case?
The transparency gap is concerning. Anthropic disclosed this to the Financial Times, but it's unclear what oversight exists. Does Congress know the details? Do Anthropic's investors? Do the engineers working on Claude's safety features know their colleagues are deploying it for intelligence operations?
National security and AI safety exist in tension. The same capabilities that make AI useful for defense make it dangerous for offense. A system that can detect vulnerabilities can also exploit them. A system that can analyze adversary communications can analyze anyone's communications.
This isn't unique to Anthropic. OpenAI, Google, and Microsoft all work with intelligence agencies. The difference is Anthropic has built its brand on being the responsible actor. That brand becomes harder to maintain when you're embedded in offensive cyber operations.
The technology is impressive. The question is who controls it and for what purpose. When engineers from safety-focused AI companies are deploying systems inside the NSA, the line between research and weaponization blurs.
We deserve more transparency about what these systems are actually doing and what oversight exists to prevent abuse. Right now, we're getting press releases about safety while engineers are embedded in Fort Meade.
