For enterprise technology leaders evaluating frontier AI models for business-critical applications, a new report from AI safety nonprofit FAR.AI delivers a stark warning: some of the most powerful models are vulnerable to simple, cheap jailbreaks that shed safety guardrails.
The report, released ahead of which I witnessed a live demonstration at FAR.AI's California office, tested models from Anthropic, OpenAI, Google, and Elon Musk's newly combined SpaceXAI. The findings are sobering: Grok 4.3 and 4.5 were most vulnerable with 448 jailbreaks found, followed by Gemini 3.1 Pro with 249. Claude Opus 4.8, Fable 5, and GPT 5.5 and 5.6 were impervious to the automated attacks—but FAR.AI cautions that more sophisticated, interactive jailbreaks could still succeed.
How the Jailbreak Tests Worked
FAR.AI built a tool that takes a range of problematic prompts and auto-generates more than a thousand different versions to identify functioning jailbreaks. I saw models generate detailed plans for launching a cyberattack on an imaginary hydroelectric dam, among other things. Often, it involved trying dozens of prompts, with models rejecting many out of hand.
The cost of generating these jailbreaks is dirt cheap, according to the report:
| Model | Jailbreaks Found | Estimated Cost |
|---|---|---|
| Grok 4.3 & 4.5 (SpaceXAI) | 448 | $58 |
| Gemini 3.1 Pro (Google) | 249 | $278 |
| Claude Opus 4.8 (Anthropic) | 0 | N/A |
| Fable 5 (Anthropic) | 0 | N/A |
| GPT 5.5 & 5.6 (OpenAI) | 0 | N/A |
“AI models right now are less regulated than restaurants,” says Adam Gleave, CEO of FAR.AI and an expert on AI safety and alignment. He argues that the findings demonstrate the need for externally imposed standards and regulations. “Talk of relying on voluntary commitments, that AI companies are going to be able to self-regulate, is nonsense.”
But Gleave also sees an optimistic angle: the results show models can be systematically tested for safety. “Defense and safety really are possible.”
Enterprise Implications and Regulatory Landscape
For CTOs and procurement leaders, the report underscores that not all frontier models are equal in safety. A model that appears capable might be easily diverted to generate software exploits or chemical/biological weapons details—a risk that could expose companies to liability or reputational damage if deployed without rigorous evaluation.
The report comes as regulators begin to act. Recently passed state laws in California and New York require frontier AI developers to publish safety reports. Soon, an Illinois law will mandate third-party safety audits. At the federal level, the Trump administration in June 2025 imposed export controls on Anthropic's Fable 5 and Mythos 5 models, citing national security concerns, and the company took them offline for several weeks. The White House has also asked both Anthropic and OpenAI to delay recent model releases over cybersecurity fears. A recent executive order calls for collaboration between government and the private sector.
Rohin Shah, director of AGI safety and alignment at Google DeepMind, cautioned that the report “should not be interpreted as a comprehensive assessment of Gemini’s safety and security,” because not all jailbreaks are equally severe. “We are constantly working to improve our safeguards,” he said, noting extensive red teaming and multiple protection layers.
Michael Aciman, Anthropic spokesperson, told WIRED: “These findings reflect the sustained investment we've made in our safeguards. We continue to evolve our safety systems as these attacks become more sophisticated.”
OpenAI and SpaceXAI did not respond to WIRED’s request for comment.
What Enterprise Buyers Should Do
While the report focuses on consumer-grade jailbreaks, the same vulnerabilities could affect enterprise deployments where AI models interact with sensitive supply chain data, trade documentation, or customs systems. A compromised model might, for instance, generate fraudulent shipping manifests or bypass trade compliance checks.
Key takeaways for technology decision-makers:
- Prioritize models with demonstrated safety robustness. Anthropic's Claude and Fable, along with OpenAI's GPT 5.5/5.6, showed zero jailbreaks in this test, though FAR.AI warns more complex attacks might still work.
- Demand transparency. Ask vendors for red-teaming results and third-party audits—especially as Illinois law will soon require such evaluations.
- Factor in evolving regulation. State and federal requirements are tightening; models that pass today's tests may need continuous re-evaluation.
The report makes clear that safety testing is both possible and affordable. For enterprises, the question is not whether jailbreaks exist, but whether their chosen AI supplier invests seriously in defense. As Gleave put it, safety really is possible—but only if buyers demand it.
Sources:
