No Frontier AI Lab Scores Above C+ on Security Controls

None of the five leading frontier AI labs — Anthropic, Google, Meta, OpenAI and xAI — scored above a C+ on publicly available evidence of their AI control practices. That's the headline finding of a recent assessment by the nonprofit Guidelight AI, and it frames an uncomfortable question for an industry asking the world to trust increasingly autonomous systems.

The argument comes from a security practitioner writing in an opinion piece for Unite.AI, and it's built on a simple premise borrowed from cybersecurity: you don't discover whether your controls work after something breaks. You test them, challenge them, and assume someone will eventually find the weakness you missed. As the author puts it, "That's Security 101."

So why should frontier AI get a different standard? The piece's core claim is that capability is moving faster than control, and that speed is precisely the reason stronger security is needed — not an excuse for weaker security.

What did the assessment actually find?

The C+ ceiling doesn't mean these companies are reckless, the author is careful to note. Additional controls may well exist behind closed doors. But "security doesn't run on assumptions." If enterprises and the public are asked to trust increasingly powerful systems, the author argues, they should expect evidence that the surrounding controls can withstand the same scrutiny demanded anywhere else in security engineering.

The specific weak spots are prevention and containment. Public evidence around gated actions, circuit breakers, and emergency containment remains limited, according to the piece.

What does this change for security teams?

The most practical part of the argument concerns speed, not novelty. The author doesn't believe the biggest near-term threat is AI inventing an entirely new category of cyberattack. The bigger problem: "AI can make the attacks we already understand happen much faster."

That has direct operational consequences. Research that once took hours can happen in minutes. Attack techniques can be chained together faster. Vulnerabilities can be identified and exploited before a human-centered security process finishes its first escalation call. Much of today's security infrastructure still assumes there will be time to respond — and sometimes there won't be. A response process built around tickets, meetings, approvals and manual escalation can become obsolete before anyone reaches step three.

The flip side is a demand aimed squarely at the labs: if frontier developers 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."

Intentions don't prove security

The author draws on a career in security testing to state the governing principle: "intentions don't prove security. Evidence does." In practice, that means you test, you monitor, you try to break things, and you confirm someone has the authority to stop the system when necessary.

Applied to frontier AI, the piece calls for four concrete things:

• Adversarial testing of controls, with measurable evidence they work
• Pre-authorized safeguards rather than improvised responses
• Clear limits on agent authority
• Proven ways to intervene or shut a system down before a failure spreads

Labs cannot improvise containment after an AI agent acts outside its intended boundaries, the author writes. As they train increasingly capable models, they must take responsibility for the actions those systems can take. Human judgment still matters, but the labs must establish the boundaries, test whether they hold, and remain accountable when they don't.

Trust must be earned

This isn't a call to slow AI innovation to a crawl. The author explicitly frames it as security keeping pace with capability — comparable urgency around making AI defensible, matched to the enormous resources going into making it more capable.

Independent review is the closing demand. It "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 piece ends on the practitioner's terms: "The AI industry wants our trust. As a security practitioner, I'm willing to give it. But first, show me the controls." For CISOs and risk teams evaluating AI vendors, that's likely to become the standard procurement question — and right now, on public evidence, no lab is acing it.