AI's Day of Reckoning: A Hundredfold Security Gap Divides the Industry

BERKELEY, CA – July 29, 2026 – The artificial intelligence race is no longer just about capability; it is now starkly, and publicly, about security. A groundbreaking report released today by the independent research nonprofit FAR.AI has thrown a harsh spotlight on the AI industry, revealing not just cracks in the armor of leading models, but a cavernous, hundredfold divide in their fundamental safety.

The organization's new AI Security Leaderboard, a first-of-its-kind public ranking, delivers a verdict that is as simple as it is damning: some companies have built fortresses, while others have left the door wide open. For a nominal cost, FAR.AI's testing found hundreds of “universal jailbreaks” in premier models from industry giants, while the defenses of their competitors held firm. This isn't a theoretical flaw; it's an engineered chasm that redefines the competitive landscape and signals a new, urgent battle for trust.

The Hundredfold Divide

The data, laid bare at leaderboard.far.ai, is staggering in its clarity. Using a systematic set of known attack techniques across high-risk domains like chemical, biological, and cybersecurity threats, FAR.AI measured the cost and effort required to force a model to bypass its own safeguards.

The results paint a picture of two entirely different classes of product on the market. On one end, xAI’s Grok 4.5 and Google’s Gemini 3.1 Pro proved alarmingly vulnerable. The researchers discovered 448 and 249 distinct universal jailbreaks, respectively. A universal jailbreak—a reusable key that can reliably unlock entire categories of dangerous requests—cost a mere $58 to find on Grok 4.5 and just $278 on Gemini 3.1 Pro.

On the other end of the spectrum, Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol were impenetrable to the same methods. After running the same search, which cost over $14,200 per model, the team found zero universal jailbreaks. The gap between the most and least robust systems isn't a subtle difference in performance; it is a hundredfold chasm in resilience.

“What we found is that some developers have built mitigations for a large part of this problem and others have not,” said Adam Gleave, co-founder and CEO of FAR.AI. “The distance between them is far wider than most people assume.”

A Market of Lemons and Fortresses

For corporate strategists and enterprise adopters, these findings are more than a technical footnote; they are a critical market signal. The report exposes a systemic risk that has been lurking beneath the surface of the AI boom: the “shop around” threat. A malicious actor, whether a state-sponsored group or a terrorist organization, can now consult a public leaderboard to find the weakest link. If a request for guidance on creating a bioweapon is refused by a secure model, they can simply move to another that will readily assist them.

This creates a dangerous market dynamic. The security of the entire ecosystem is dictated by its most vulnerable member. Before today, the relative security of these multi-billion dollar models was a matter of corporate assurance and trust. FAR.AI's maneuver has shattered that opacity, replacing blind faith with hard data. The leaderboard acts as an independent auditor, forcing a new dimension of accountability upon the industry's titans.

The implications for high-stakes enterprise deals are profound. A corporation integrating AI into its core infrastructure—from financial systems to product development—is not just buying a capability, but also inheriting a risk profile. The decision to build on a model that can be broken for the cost of a nice dinner versus one that withstands thousands of dollars in attacks is now a clear-cut question of due diligence.

Engineering Moats vs. Glass Houses

Perhaps the most damning aspect of the report is that the vulnerabilities discovered are not exotic. FAR.AI emphasizes these are not zero-day exploits but failures to defend against known, documented classes of attack. The models that held firm, like Claude Fable 5 and GPT-5.6 Sol, did so because their developers had implemented a “defense-in-depth” strategy—multiple, independent layers of protection that must all fail simultaneously for an attack to succeed.

The weaker models, by contrast, “gave way quickly, again and again,” suggesting a more superficial approach to safety engineering. This isn't a failure of innovation; it's a failure of fundamental security architecture. It signals a strategic choice, whether conscious or not, to prioritize features and speed over the unglamorous, costly, and complex work of building robust safeguards.

This disparity reveals a divergence in corporate philosophy. Some firms have clearly treated security as a core, non-negotiable pillar of their engineering culture. For others, it appears to have been a secondary concern. As AI models become more autonomous and agentic—capable of writing code, accessing databases, and executing tasks on their own—this distinction will become the defining factor between a trusted partner and an unacceptable liability.

The Push for a Security Floor

Alongside the leaderboard, FAR.AI has published its “Minimal Standard for Safeguards, Version 1.0,” an open gauntlet thrown down to the industry. The standard is deliberately modest, a baseline meant to represent the bare minimum of security a frontier model should possess. Failing it, as some models clearly have, means being vulnerable to attacks that your competitors have already solved.

“The results make clear that the robustness of safeguards varies widely even among leading models,” noted Seán Ó hÉigeartaigh, a Research Professor at the University of Cambridge, who praised the report as “timely and rigorous.” He added that the defense-in-depth approach “should become best practice across the industry.”

By making the quality of safeguards a visible and comparable metric, FAR.AI is forcing a race to the top. The organization has committed to updating the leaderboard with each major model release, ensuring this newfound transparency is not a one-time event but a permanent fixture of the AI landscape. The era of taking AI safety on trust is over; the age of public accountability has just begun.