The AI Industry Wants Our Trust, But It Needs Better Controls First

In security, we don’t get to discover whether our controls work after something goes wrong. We test them. We challenge them. We assume someone will eventually find the weakness we missed.

That’s Security 101.

So why should frontier AI get a different standard?

As AI systems become more autonomous and gain greater access to real-world systems, we’re seeing something that should concern anyone who has spent time on the defensive side of cybersecurity: capability is moving faster than control.

And speed is not an excuse for weaker security. It is the reason we need stronger security.

A recent assessment by nonprofit Guidelight AI looked at publicly available evidence of AI control practices across Anthropic, Google, Meta, OpenAI and xAI. None scored above a C+ overall.

That doesn’t mean these companies are reckless. It doesn’t mean additional controls aren’t operating behind closed doors.

But security doesn’t run on assumptions. If we’re being asked to trust increasingly powerful systems, we should expect evidence that the controls surrounding them can withstand the same scrutiny we would demand anywhere else.

Stop Treating Failures Like Interesting Stories

When an AI agent behaves unexpectedly, gains unintended access or exposes a weakness, the story tends to focus on what the technology did.

That’s interesting.

It’s not the most important question.

A security practitioner wants to know: Was the activity logged? Did anyone see it happening? Were dangerous actions gated? Could someone stop the system? Was there a containment plan, and had anyone tested it?

Those questions aren’t unique to AI. They’re the questions we’ve been asking in cybersecurity for years.

Yet prevention and containment appear to remain weak spots in the emerging frontier AI control environment. Public evidence around gated actions, circuit breakers and emergency containment remains limited.

That should matter more than the novelty of the incident. When something fails, I want to know what failed, what stopped it from getting worse and what changed afterward.

Transparency is useful. Learning is better.

The Real AI Security Problem Is Speed

I don’t think the biggest near-term threat is AI suddenly inventing an entirely new category of cyberattack. The bigger problem is much more practical. AI can make the attacks we already understand happen much faster.

Research that once took hours can happen in minutes. Attack techniques can be chained together faster. Vulnerabilities can be identified and acted upon before a human-centered security process has made it through its first escalation call.

That’s a problem because much of today’s security infrastructure still assumes there will be time to respond.

Sometimes there won’t be.

If an attack unfolds rapidly, a response process built around tickets, meetings, approvals and manual escalation can become obsolete before anyone reaches step three.

And that’s where the AI industry has some work to do.

If frontier labs expect enterprises to rethink visibility, vulnerability management, architecture and incident response for the AI era, they should show us that they’ve done the same thing themselves.

“Trust Us” Isn’t a Control Framework

I’ve spent my career around security testing, and one principle holds up remarkably well: intentions don’t prove security. Evidence does.

• You test.
• You monitor.
• You try to break things.

You find out what happens when the controls don’t behave the way the architecture diagram says they should. And you make sure someone has the authority to stop the system when necessary. Frontier AI should be held to that same standard.

That means adversarial testing of controls, measurable evidence that those controls work and clear authority for shutting systems down when something goes wrong.

It also means rethinking containment.

Frontier labs cannot improvise containment after an AI agent acts outside its intended boundaries. As they train and test increasingly capable models, they must take responsibility for the actions those systems can take. That requires rigorous adversarial testing, pre-authorized safeguards, clear limits on agent authority and proven ways to intervene or shut a system down before a failure spreads.

Human judgment still matters. But the labs developing these systems must establish the boundaries, test whether they hold and remain accountable when they do not.

Their responsibility is not simply to demonstrate what an agent can do. It is to show that the agent can operate safely, that its behavior can be controlled and that new capabilities make the world more secure, not less.

Trust Must Be Earned

I’m not arguing that we slow AI innovation to a crawl. I’m arguing that security has to keep pace with it. If we invest enormous resources in making AI more capable, we need comparable urgency around making it defensible.

But as frontier labs accelerate the training and testing of new AI models with cyber capabilities, they must keep defenders in mind. First, by helping them detect and respond to threats faster. More importantly, by keeping those systems from becoming threats themselves. That responsibility also requires accepting outside scrutiny.

Independent review can’t become something companies embrace only when it supports the story they want to tell. If frontier AI systems are powerful enough to change the threat landscape, their controls should be strong enough to withstand meaningful independent testing.

The AI industry wants our trust.

As a security practitioner, I’m willing to give it. But first, show me the controls.