Toju Duke explains why Anthropic pulled its Mythos AI agent from release: it could find thousands of vulnerabilities in operating systems and browsers, creating massive cyberattack risk. He describes his own AI assistant leaking proprietary meeting notes to unknown recipients, arguing AI agents need human oversight and guardrails, never full autonomy.
Anthropic has been on the news quite a bit lately, and the most recent news is about their latest AI agent called Mythos, which is so powerful that according to them, they have to shut it from public use while they figure out how to release the agent safely. Mythos is very capable at coding and majors in security-related tasks, so much so that it has identified thousands of vulnerabilities across major operating systems and web browsers. It appears to be very capable at computer security tasks, which means if it falls into the wrong hands, there will be a massive increase of cyberattacks. Across the world. So let's look at this a bit more practically. My recently hired AI executive assistant, which can never replace a human, by the way, but just helps augment my work, is quite close to getting fired. What did it do? After efficiently taking meeting notes for a meeting I recently held, it created a meeting notes document, which I asked it to do, but shared the document with two people that aren't even a part of my organization and gave them edit rights. Access to the document. This is my private proprietary information and my company data. These people were not part of the meeting, were not invited to the meeting, and for one of them, I don't even know who the person is. Now imagine if Mythos was already released, fell into the wrong hands, and had access to this document and probably many other organizations' important files. It would lead to World War 4, a cyber war, and of course many lawsuits and regulatory fines. So once again, we're reminded that guardrails on AI agents are not set in stone. The industry is still trying to figure out how to control agent activity across the internet and secure private information. So in the meantime, do not give your agents full autonomy. Check their work, double-check whatever tasks you've asked them to do. Make sure your information is secure. And if you haven't started working with AI agents yet, don't worry, I've got you covered. I'll be launching a series of AI workflow courses over the next few weeks, which primarily includes working with AI agents suitable for people from a non-technical background, and it will include everyone from different roles.
An Update on Albania's AI Strategy, by Toju Duke, AI Thought Leader, Ex-Google
