Notes From Underground (#21): Bringing on Bugmageddon
A New AI Frontier And the Overlooked Dangers of Selective Enforcement
A belated happy Juneteenth everyone! And in other news…
11 days ago, the Trump Regime effectively forced two of Anthropic’s newest AI cybersecurity systems offline.
To be sure, though, the Trump Regime’s reasons for doing so may seem to make sense—at least initially. The two Anthropic models, “Mythos 5” and “Fable 5,” could indeed raise major national security concerns.
For example, Anthropic and its cybersecurity experts “spent months describing Mythos,” the earlier version of these two systems, “as one of the most capable—and most dangerous—AI models ever built.” That in itself could raise justifiable alarms.
Anthropic researcher Nicholas Carlini even demonstrated just “how easy it had become to use new models to break into [web] systems.” Most notably, Carlini showed how Mythos found many weaknesses in even “the most battle-tested [piece] of software.” Even when Carlini himself, a cybersecurity expert, had never been able to spot them before.
Again, this could definitely be seen as a problem worthy of government action. After all, major sections of the U.S. economy run on software products like the one that Mythos threatens to hack into if placed in the wrong hands. Banks especially worry that software like Mythos—even with greater safety checks—could expose huge weaknesses in software that keeps our financial system running.
That’s not even to mention the accompanying concerns about what this means for the privacy of an everyday person on the web.
I repeat, pretty simple logic, right? Here’s the problem: there is another AI cybersecurity system. Its ability to discover weaknesses in software tops Anthropic’s banned models. And yet…
Why? Well, it could be because it’s an OpenAI model. It could be because OpenAI never criticized or pushed back against the Trump Regime for threatening to misuse its AI technology at the Pentagon. It could be because OpenAI and Sam Altman never filed a lawsuit against the Trump Regime. Remember that Anthropic’s legal actions against the Regime are still ongoing.
Consistent with everything else this Administration does, WSJ reporting suggests that “[the Trump Regime] told Anthropic it had 90 minutes to pull down the model and didn’t provide details about the security risk,” per a source close to the company.
In short, I am not sure that the Trump Regime’s approach to Anthropic was merely miscommunication on one or both sides. To the contrary, I think it may reflect Trump’s attempt to control the technology and who has the power to use it.
If true, this should concern all of us. Yes, I admit, all of the AI companies, including Anthropic, have made some ethical missteps in their use of AI. However, any missteps like the kind Anthropic made are nothing compared to the danger of surrendering AI technology and its effective oversight to a Regime obsessed with control and punishment of those who dare to oppose it.
This technology has the power to completely change the security of the web, the world, and the privacy of anyone in both. If we leave this technology with the wrong people—government or private actors—it could create an unheard-of danger to the freedoms that we hold dear.
Stay tuned. And stay attentive. Big Brother may be watching too.