Senate testimony claims Anthropic’s Mythos AI breached NSA and Cyber Command systems in hours, prompting a U.S.-ordered shutdown. On June 12, the Trump administration directed Anthropic to restrict access to Fable 5 and Mythos 5, its two most capable models, exclusively to US citizens. Because verifying every user’s nationality in real time isn’t practically possible, […]
Does it say more about the AI or about the NSA systems? What genius lets AI anywhere near these systems?
The discourse about AI across Lemmy is so interesting. People heavily deny the capability of current AI systems, the nearly exponential growth in it’s capability since 2017, that we are in a race to AGI / ASI and the winner (assuming we solve the control / alignment problem) will essentially have a perpetual leverage over everyone else. Most I think are just seeing AI videos and images and think it is a stupid and useless toy poisoning their FYP.
Like I’ve had people literally call me a cultist or insane or brainwashed when I’m trying to give some caution that shit is getting serious and then you get articles like this.
I’ll also note that cybersecurity hygiene, including simple things like patching systems in a timely manner, is notoriously bad on government systems because cybersecurity is not well funded. DOGE’s cuts didn’t help either. An AI agent that exploits known security vulnerabilities like that isn’t as special or surprising as it might sound at first blush.
It was also reported that Trump ordered the restrictions because one of the Anthropic employees was a Democrat. At this point, I’m not sure what to believe.
I would think China would be more of a concern for the US in this as they are the only real contender. Maybe he was concerned someone not loyal may come across some unredacted files and pull a Snowden?
Without getting in to the politics or morals of AI, my first question to those who make statements like “it’s just fancy autocomplete”, or “AIs always hallucinate” is when they last used an AI tool, and which model(s) they used. The answer is often that they tried ChatGPT when it came out, thought it was a neat toy, and haven’t experimented much with it since then.
Most regular people aren’t paying for AI tools yet, and don’t have professions that bring them into regular direct contact with frontier AI models. They’re far from from perfect, and not AGI/ASI, but family and friends have been astonished when I’ve shown them Fable 5, GPT-5.5 doing things they understand with long-running agentic actions.