Mythos: Why is it time to worry about AI

By now, you might have heard about Anthropic’s new model, Mythos. If not, the one-liner version is that Anthropic announced its most powerful model called Mythos last week; however decided not to release it to the public since it is too powerful to handle.

In their testing, the company found out that Mythos is capable of finding and exploiting security vulnerabilities in every major operating system and web browser. Things which humans could never detect. It even found 27 year-old bug in OpenBSD — one of the most secure operating systems.

Reportedly, Mythos performed 94.6% in GPQA tests (vs. Claude Opus 4.6 at 91.3%, GPT-5.4 at 94.3%). In simple terms, this is a standard test whether AI can answer PhD-level questions in biology, physics, and chemistry — the kind that would stump most experts!

Why worry?

Because Mythos is not just about software security. It is a sign of a much bigger problem a lot of AI experts (including Geoffery Hinton, the “godfather of AI”), has been warning us all along. This could be that threshold of advancement where AI could be beginning to affect trust, public safety, politics, and the stability of everyday life.

Though Anthropic made the right call to not release the model to public, it is just a matter of time something similar gets into public hands. If not by Anthropic, then by someone else — putting this kind of power into the hands of a disgruntled employee, a bad state actor, or just a bored teenager. I want you to look at these real cases of AI misuse and then imagine an almost equitable ability for anyone to do the same. One just need to have the intent:

1. Cybersecurity — Arup deepfake scam, where a fake CFO video tricked a finance worker into a $25M transfer.

2. Biological weapons — In an experiment, AI researchers designed toxin variants that passed DNA vendor safety checks, evading bioweapon detection. One can imagine the catastrophe of this getting in wrong hands.

3. Misinformation — Facebook’s algorithms amplified hate against Rohingya in Myanmar, fuelling ethnic violence and genocide (Harari, Nexus). Similarly, AI-generated videos of Muslim women complaining about Australian way of life, went viral, sparking outrage before glitches exposed them as fakes created from offshore accounts.

4. Social engineering — The FBI’s Internet Crime Complaint Center (IC3) reported that AI chat tools scaled romance scams to $19M in losses last year.

5. Autonomous AI — Flaws in Moltbook, an AI agent social network (which got viral recently), exposed 35,000 emails and API keys to hackers. Attackers used machine learning to infiltrate software updates, compromising entire networks downstream.

What do we need to do?

There is no easy answer to this question. The answer is neither in stopping AI advancement, since that is practically not possible. Neither is it in having a blind optimism. Partly the answer lies in what humans have always been great at — resilience!

First, we need stronger verification systems, better labelling of synthetic content, and clearer accountability for platforms and deployers. This model of “I build, you regulate” is too fragile. No government or regulator in the world has the kind of funding these tech firms have.

Second, digital literacy (or rather cyber awareness) needs to be treated as an immediate need in our educational institutions. It is not an option or can be left for society to figure out on its own.

Third, governments need to treat AI as a governance issue, not just a competitive advantage or economic issue, with rules that match the scale of the risk.

The Mythos story captures a turning point. AI is no longer something we only use to do things faster. It is becoming something that can shape what people believe, how economies grow, and how politics is conducted.

The question is no longer whether AI will change society. It already is. The real question is whether society will build the institutions strong enough to live with it.