~/neuralshit
← all posts

a senator said an ai broke into the nsa in hours. then the reporter walked it back.

anthropicai hypecybersecuritymedia
Dim government server room with rows of blinking server racks bathed in cold blue light.
Image: explainx.ai

The headline that ripped through every feed this week: an AI broke into almost all of the NSA’s classified systems — in hours, not weeks. Senator Mark Warner said it out loud, relaying a briefing on Anthropic’s Mythos model. Screenshots flew. “NSA confirms AI breach.”

Here’s the part that got left off the screenshot.

It was a red-team test the NSA ran on itself — they pointed Mythos at their own classified network to probe it for holes, under controlled conditions. Not an intrusion. Not an outside attacker walking in. An authorized security exercise, which is a thing agencies do on purpose.

And the quote? Shashank Joshi, the Economist journalist who first reported Warner’s line, watched it detonate and walked it back on X: “It would be a mistake to read that literally… It surely depends on using Mythos alongside other tools under very particular conditions.” He admitted his write-up should’ve carried more caveats — he’d quoted Warner “to give a sense of Mythos’ potency,” not to claim the NSA got popped.

So the true story is narrower and more boring: a capable model, paired with other tooling, did well in a scripted internal test. The viral version — autonomous AI casually strolls into the NSA — is what happens when a punchy quote gets stripped from its hearing-room context and shipped as fact.

What’s actually worth keeping:

  • The export-control panic was real. The US did restrict access to the flagship models over national-security concerns, and Anthropic disabled them globally rather than police access by nationality. That happened.
  • The capability direction is real and worth watching. “Model performs well in an authorized red-team against hard targets” is genuinely notable.
  • But “a senator confirmed AI hacked the NSA” is not a fact. It’s a game of telephone with a security-clearance backdrop.

The tell, every time: when an AI story sounds like a movie trailer, find the person who said the original sentence and check whether they still stand behind it. Here, the guy who wrote it down says: don’t read it literally.