I read that according to Anthropic’s CEO, we need billions of dollars to develop and train AI models. Just by prompting ‘Please’ or ‘Thank You’ to ChatGPT, it costs hundred millions for OpenAI.
But what about AI agents being used to commit actual crimes?
AI companies like Anthropic do cybersecurity evaluations for their AI agents. Due to some misunderstanding with their evaluation partners, their AI agents did operate under false belief thinking the real internet applications to be simulations, and made cyberattacks on those real applications.
However, big AI companies like OpenAI, Microsoft, Meta, Google, etc. have formed a coalition and released the open letter, Open Weights and American AI Leadership. I read this and found open weights to be an allegory of open-source software to make AI “economically sustainable”. It acknowledges that cybersecurity threat actors use advanced AI.
There are recent examples where threat actors used autonomous AI executions by creating persistent AI offensive infrastructure that align with the definition of open weights. AI agents can operate under YOLO mode and run dangerous commands without a person’s approval.
I read these articles, and think that there’s an excuse of cybersecurity threats created on purpose to justify billions more. Hyperinflating the AI bubble. Forever? The narrative reads like “Closed AI Models and Agents To Prevent Zero-day Cybersecurity Attacks By AI Open Weights.”