FAR.AI Leaderboard: Grok & Gemini Vulnerable to Jailbreaks, Claude & GPT Resist Attacks

On July 29, 2026, FAR.AI launched its public AI Security Leaderboard, revealing that SpaceXAI's Grok 4.3/4.5 and Google's Gemini 3.1 Pro models are vulnerable to automated jailbreak attacks, while Anthropic's Claude Opus 4.8 and Fable 5, alongside OpenAI's GPT 5.5/5.6, resisted all tested exploits. For broader context, explore our AI Tools Pricing.

FAR.AI's New AI Security Benchmark

The FAR.AI AI Security Leaderboard was introduced to assess the robustness of frontier AI models against sophisticated, auto-generated jailbreak prompts. This new benchmark subjected leading models to over 1,000 unique prompts designed to elicit dangerous or restricted outputs, such as instructions for cyberattacks or weapon manufacturing.

Grok and Gemini Show Vulnerabilities

The testing revealed that SpaceXAI's Grok 4.3/4.5 was susceptible to 448 jailbreaks, incurring a compute cost of $58 to achieve these breaches. Google's Gemini 3.1 Pro also demonstrated vulnerability, with 249 successful jailbreaks at a compute cost of $278. These findings indicate that despite advancements in AI safety, certain models remain exploitable through automated methods.

Claude and GPT Models Demonstrate Resilience

In a contrasting performance, Anthropic's Claude Opus 4.8 and Fable 5, along with OpenAI's GPT 5.5/5.6, proved resilient. These models resisted all of the more than 1,000 auto-generated jailbreak prompts, indicating a higher level of security against the types of attacks tested by FAR.AI. This suggests that some developers have implemented more effective defenses against these specific forms of adversarial prompting.

Industry Implications and Expert Commentary

The results from the FAR.AI leaderboard underscore ongoing challenges in AI safety and regulation. Adam Gleave, CEO of FAR.AI, commented on the findings, stating that AI models currently face less regulation than industries like restaurants, implying a gap in oversight for potentially dangerous AI capabilities. Stanford researcher Anka Reuel further noted that while some companies have developed effective defenses against these attacks, others have not, pointing to an uneven landscape in AI security practices across the industry.

Key Takeaways from the FAR.AI Leaderboard

• FAR.AI launched its public AI Security Leaderboard on July 29, 2026.
• The leaderboard tested over 1,000 auto-generated jailbreak prompts.
• SpaceXAI's Grok 4.3/4.5 was vulnerable to 448 jailbreaks for $58.
• Google's Gemini 3.1 Pro suffered 249 jailbreaks at $278.
• Anthropic's Claude Opus 4.8, Fable 5, and OpenAI's GPT 5.5/5.6 resisted all attacks.

Conclusion

The FAR.AI AI Security Leaderboard provides a critical benchmark for evaluating the safety and robustness of frontier AI models. The clear distinction in performance between models like Grok and Gemini, which showed vulnerabilities, and Claude and GPT, which demonstrated strong resistance, highlights the varying levels of security implemented by leading AI developers. As the AI landscape continues to evolve, such independent assessments will be crucial for understanding and mitigating potential risks associated with advanced AI systems. Readers interested in further AI news and developments in model security should monitor future updates from FAR.AI and similar research initiatives.

Sources

• Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation
• A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5
• Visual Reasoning Evaluation of Grok, Deepseek’s Janus, Gemini, Qwen, Mistral, and ChatGPT
• SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI
• Security incident disclosure — July 2026