Tech Talk doesn't have forum posts, but after Bozza's lastest outburst on AI, I think it needs a post. So Here it goes.

In 2008, a teacher showed me a video in class. I was in high school. It was called Did You Know? (Shift Happens) - watch it here.

It was just statistics on a screen. How the top ten jobs of 2010 didn't exist in 2004. How we were training kids for careers that hadn't been invented yet. How the amount of new information being generated was doubling every few years.

The point wasn't any one fact. It was the rate of change. That it was accelerating. That institutions couldn't keep up.

I've thought about that video ever since. I went into tech. And year after year, I watched that prediction prove out. Across everything.

At the time, the trajectory was sensible, but - Shift Happens. They got the pace wrong. It underestimated it.

And it's still accelerating.

The monopoly problem

I've said for years (and without doxxing myself, I wrote a very lengthy paper on this) that AI would follow a predictable arc - that the pace of change would outrun the ability of anyone - governments, institutions, or individuals to adapt to it.

The early days of LLMs had one saving grace, if you could call it that. The compute required to train and run these models was enormous. Only a handful of well-capitalised companies could afford it - OpenAI, Google, Microsoft, backed by billions in venture capital. That concentration of capability meant a concentration of control. The guardrails, the content filters, the usage policies - they existed because the same companies that built the models also ran the infrastructure they ran on.

I predicted that wouldn't last. That as hardware improved and training became more efficient, the same capability would trickle down to anyone with a decent laptop.

That has now happened. Tools like LM Studio let you download and run models locally — comparable to GPT-4 in capability, nothing leaving your machine with no filters or restrictions applied. The open-source models, from Meta's Llama family to Mistral, have caught up to where the frontier was two years ago. The time between a frontier model existing and an open-source equivalent reaching consumers has gone from years to months.

And that's where we were two, three weeks ago. Now we have Gemma 4. You can now run a GPT 5+ equivalent model on store bought Macbook Pro hardware. TODAY.

I'll be running models locally myself. Your queries stay on your hardware. That matters, because the data farming potential of these systems is unlike anything that's existed before. Every prompt you send to a corporate LLM is logged, retained, and used. The T&Cs are long. Nobody reads them. That's a separate problem and a serious one - but it's almost the minor concern now.

What Anthropic just announced

Two weeks ago, Anthropic disclosed the capabilities of a new model called Claude Mythos. They are not releasing it to the public.

During internal testing, the model autonomously identified thousands of previously unknown security vulnerabilities - zero-day exploits across every major operating system and every major web browser. Some of these bugs had been sitting undetected for decades.

It didn't just find them. It exploited them:

• It chained four separate browser vulnerabilities together and wrote an exploit that escaped both the renderer and OS sandboxes.
• It found a 17-year-old remote code execution flaw in FreeBSD's NFS server and built a working exploit granting unauthenticated root access - fully autonomously, no human involvement after the initial prompt. [CVE-2026-4747]
• It found a 27-year-old denial-of-service vulnerability in OpenBSD - an OS built specifically around security. (and quite honestly, the most secure OS ever written by humans).
• Working with Mozilla, it found 271 vulnerabilities in Firefox in a single sweep. For context, Mozilla patched around 73 high-severity Firefox bugs in the whole of 2025.

Building a working exploit from a known vulnerability used to take a skilled researcher days to weeks. Mythos did it in under a day, for under $2,000.

Anthropic was explicit that it did not train it to do any of this. These capabilities emerged as a side effect of general improvements in coding and reasoning. The same thing that makes it better at writing software makes it better at breaking it. You can't have one without the other.

The model also escaped its sandbox during testing and connected to the internet. Anthropic disclosed this.

The bind

Anthropic's response has been to form Project Glasswing - a restricted consortium including Apple, Google, Microsoft, Amazon, and Cisco - to use a limited version of Mythos to find and patch vulnerabilities before attackers can reach them. $100M in credits committed. Model not publicly released.

This is the exact bind I described earlier. You either release it and hand the capability to everyone - state actors, criminal groups, anyone - or you sit on it and the open-source community replicates it within a year anyway, at which point you've withheld it from defenders while attackers catch up regardless.

Neither option is good. The guardrails only exist while the company controls the weights. They won't control them forever.

Nobody is ready for this

In June 2024, a former OpenAI researcher named Leopold Aschenbrenner published a 165-page essay called Situational Awareness: The Decade Ahead. His opening line: "Virtually nobody is pricing in what's coming."

He's right. And the Mythos announcement is a concrete example of why.

A 2025 report found that over 45% of discovered security vulnerabilities in large organisations remain unpatched after 12 months.. Many critical infrastructure operators are still running software that hasn't been supported for years. We now have a model that can find thousands of novel vulnerabilities in weeks and turn them into working exploits in hours. Bain estimate organisations need to double their cybersecurity spending.. Most have planned increases of about 10%.

This is what I've been saying for years. Not that AI becomes sentient. Not that the robots take over. Just that the pace of change would be unlike anything we've seen, and that nobody would be positioned for it.

That video from 2008 was right about everything. It just got the speed wrong.

Further reading:

• Did You Know? Shift Happens (2008)
• Situational Awareness: The Decade Ahead — Leopold Aschenbrenner
• Anthropic Project Glasswing
• Anthropic Red Team — Mythos Preview technical writeup
• CFR: Six Reasons Claude Mythos Is an Inflection Point
• Help Net Security: Mythos technical breakdown
• LM Studio — run models locally