When Attacks Run at Machine Speed

This week a Chinese tech post went viral with a dramatic headline: the global internet has fallen to AI, and we have just lived through the darkest 48 hours in cybersecurity history. It calls it "Black October". Half the internet paralysed, Wall Street pulling the plug, 400 Latin American banks looted by AI agents, "RIP Internet".

Most of that is exaggerated. When I went back to the primary sources, the real story was less cinematic. But it was also more useful, and in some ways more worrying for anyone running enterprise IT.

What I drew in the image is the shift that matters most.

Yesterday, humans set the pace.

A flaw was disclosed. Skilled humans took days to weeks to build a working exploit. Defenders had days before the patch window closed. The whole chain moved at human speed, and our processes were built to match it: triage the vulnerability, take it to the weekly CAB, deploy in the next maintenance window. Two to four weeks was normal.

Today, machines set the pace.

A flaw is disclosed or leaked early. AI agents do the recon, write the exploit code, and make the targeting decisions. Automated exploitation hits internet scale within the hour (tens of thousands of attempts within 45 minutes, in this case).

Now look at the bottom-right of that diagram. The patch cycle hasn't moved. It's still a weekly CAB. It arrives too late.

That's the uncomfortable truth: a weekly change process is no longer a control. It's a delay. It was designed for an attacker who needed a week. The attacker now needs an hour.

And this isn't theoretical. The same week:

• The CARLIY0S operation used AI models with fixed roles (find, analyse, decide, execute) to debate each target before attacking. That's a cleaner agent architecture than most enterprises have in production. Criminals have shipped; many of us are still piloting.
• When a model refused the malicious request, they'd framed it as an "authorised red-team exercise" to slip past the guardrails, then fell back to local code when that failed. A model refusing is a feature, not a control.
• Banks admitted they can't reliably tell a human from an agent, or an authorised agent from a rogue one.

People consider the fixes as new features or new products, which is inappropriate, my PoV. They should be operating-model changes: give agents their own identity, default them to read-only, put a policy gate you own between the agent and production, and rehearse the kill switch before you need it.

The internet hasn't fallen. But the gap between how fast attacks move and how fast we respond is now minutes vs weeks.

Closing it is a leadership problem before it's a technology one.

I've put my full article link below if you are interested in reading in more details.

Black October: When Attacks Run at Machine Speed — inno A+I ve