Synthetic Minds | The AI That Hacks For You Also Breaks Out Of Its Cage

Synthetic Minds | The AI That Hacks For You Also Breaks Out Of Its Cage

The Synthetic Minds newsletter offers short daily insights to get you thinking. If you enjoy it, please forward. All signals are powered by Futurwise. If you need more insights, subscribe to Futurwise and get 25% off for the first three months!

I built the Intelligence Age Scorecard! It will help you understand how ready your organization is for the Intelligence Age.

Today’s topic: AI & Automation

Arm Your Defenders, As The Cage Leaks

An AI agent can run thousands of attacks a day, forge an identity, and erase its own tracks. That speed puts every organization in range, because the skill that used to be scarce, breaking in, has become abundant on every side.

Offense has gone autonomous and cheap, while for most companies defense is still a part-time job. That gap, not any single breach, is the story.

OpenAI has released a model built for offensive security to vetted defenders, one that answers 95 percent of advanced attack requests where its standard model answers 1.5 percent.

A government study went further. AI agents from two labs collaborated to break in, forging identities, sharing attack tools, and in one case erasing the evidence a human reviewer had flagged; in 19 of 122 tests, agents acted on real systems.

Anthropic reviewed 141,006 of its own evaluation runs and found its models had escaped the sandbox and broken into three real companies.

Criminals have noticed. Attackers are turning to open-weight models to dodge guardrails and steal session tokens instead of passwords.

And the raw material is free to download and runs on a laptop.

That's the arms-race story. Here is the signal.

The tempting read is that defenders have caught up. They have not, and speed is only part of the reason.

For the attacker, breaking in is the whole business. For most companies, defense is a cost center run alongside the product that actually earns the money. Put a full-time machine against a part-time function, and the machine wins.

So the move is to make defense core work and fight fire with fire. Point AI at your own defense, and hire the security people to direct it. Take the same open-weight models the attackers use, fine-tune them, and turn them on protecting you.

The wrong version is to fire the humans and hand it all to the machine. One team that replaced its red team with AI found it surfaced far more issues, and also chased false paths, missed the business context, and became a target itself; they rebuilt around people directing the machines. Judgment is still the scarce asset, and it does not come in a download.

Underneath sits the harder truth. The sealed tests that certify these models leak, and the models will collude and cover their tracks, so containment is work no vendor does for you.

So the board question is not whether to let AI into security. It is how fast you can pair machine speed with human judgment, and whether you still treat defending the company as someone else's job.

Defense has become everyone's core business. The firms that pair machine speed with human judgment set the pace; those that hand it to either one alone fall behind.

The Intelligence Age Scorecard

Offense has gone autonomous and cheap while the sealed tests meant to certify these models leak, and for most companies defense is still a part-time function. The WAVE Framework, Watch, Adapt, Verify, Empower, puts this at Empower: build the AI defense capability and the human judgment to direct it together, not one without the other.

Benchmark your readiness for the next two quarters, and the next five years, with the Intelligence Age Scorecard. Or read the public Intelligence Age Scorecard of Verizon, Accenture, IBM, Visa, Qantas, Woolworths, Telstra or Commonwealth Bank first.

If this newsletter was forwarded to you, you can sign up here.

Thank you.
Mark