Misanthropic: on Mythos, bad human behaviors and systems vulnerabilities | Michael Cembalest | JP Morgan Asset Management

Full Report: Link

Wrapping up

"An optimist might say, “the good news is that there’s a finite number of important vulnerabilities which will now eventually be fixed, after which the online ecosystem will be in better shape; and it’s great that an AI-safety focused lab like Anthropic is the one that has built such a powerful model first”. That said, Anthropic’s lead researchers indicate how hard of a road this may be: “Working with this model has been a wild ride. We’ve come a long way on safety but we still expect the next capability jump of this scale to be a huge challenge” [Anthropic’s Sam Bowman]. Anthropic also conceded that their current control and monitoring methods could be inadequate to prevent catastrophic misaligned action in more advanced future systems."

"Cyberhacking could be the ideal domain for frontier AI given the steep “gradient descent” involved: high returns to reasoning, the immediate ability to see if it works and lots of training data. And as shown below, 2025 had already set a record for published system vulnerabilities according to the Common Vulnerabilities and Exposures database maintained by the US Department of Homeland Security."