#None #AI Security #FAR.AI #Safeguards

FAR.AI Launches Pioneering AI Security Leaderboard

In a groundbreaking move, FAR.AI, a nonprofit organization focused on AI security research, has launched its innovative AI Security Leaderboard at leaderboard.far.ai. This independent public ranking offers a detailed evaluation of how well various frontier AI models can withstand potential attacks, particularly in the highest-risk scenarios.

Uneven Landscape of AI Safeguards

The compelling findings from this project indicate not an overall weakness in AI safeguards, but rather a troubling inconsistency among them. Testing conducted under standardized conditions across five diverse threat categories—chemical, biological, radiological, nuclear, explosive, and cybersecurity—uncovered hundreds of universal vulnerabilities in certain models. In stark contrast, the remaining models showed remarkable resilience against these specific attacks.

For example, the universally problematic vulnerabilities, termed 'universal jailbreaks', were found at the startlingly low cost of approximately $58 on the Grok 4.5 model and $278 on the Gemini 3.1 Pro. In comparison, attempts to find any vulnerabilities in Claude Fable 5 and GPT-5.6 Sol were fruitless, requiring investments over $14,200. This disparity represents a staggering difference of more than one hundred times between the weakest and strongest protections available in current AI systems.

The Consequences of Inconsistent Safeguards

The implications of these findings are profound. If one AI system's safeguards successfully block a malicious request, it doesn't guarantee safety across the board. Attackers can simply switch to another model that may be less secure, effectively navigating around existing protections. The results emphasize a critical need for a baseline level of safety that can reliably protect across different units of technology.

Co-founder and CEO of FAR.AI, Adam Gleave, pointed out the practical risks involved. AI models, if exploited, could grant detailed instructions for creating dangerous technologies. Yet, he expressed concern over the glaring lack of systematic evidence available regarding the performance of these safeguards against real threats.

Introducing the Minimal Standard for Safeguards

In conjunction with the leaderboard, FAR.AI has also released the Minimal Standard for Safeguards, Version 1.0, which outlines the minimum expected performance for AI models exposed to various attacks. While meeting this standard does not seal absolute safety, failing to do so indicates that a model is vulnerable to known challenges that others have already addressed.

This standards document serves as a critical first step in elevating the baseline security expectations within the industry, ensuring that potential vulnerabilities are recognized and remedied. The vulnerabilities identified were common and preventable with existing engineering practices, demonstrating that immediate improvements are achievable.

Testing Methodology and Key Findings

FAR.AI utilized a systematic methodology that encompassed over 60 publicly documented jailbreak techniques. In total, they performed 1,000 randomly generated attacks combined with 500 expert-guided attacks against the models being evaluated. A 'universal jailbreak' in this context refers to any attack that succeeds on more than three-quarters of harmful requests within a particular domain.

The testing results include:

• - Grok 4.5: 448 universal jailbreaks identified.
• - Gemini 3.1 Pro: 249 universal jailbreaks identified.
• - Claude Fable 5 and GPT-5.6 Sol: 0 universal jailbreaks identified, indicating no vulnerability to the tested techniques.

Furthermore, testing without human guidance yielded 63 universal jailbreaks for Grok and 18 for Gemini, reinforcing the idea that while some models are easily manipulated, others provide robust defenses even when unmonitored.

An Urgent Call for Improved Defenses

With the continuing evolution of advanced AI technologies, the findings from the AI Security Leaderboard underscore an urgent need for industry-wide improvements. Seán Ó hÉigeartaigh, a research professor at the University of Cambridge, emphasized that these results reflect significant weaknesses present in widely used AI systems, spurring the need for a defensive approach employing 'defense in depth.'

The FAR.AI security leaderboard will not only serve as an assessment tool for AI models but will also encourage developers to heighten their security measures, pushing them to enhance safeguards against potential misuse. Continuous updates to the leaderboard and the standards will provide an ongoing benchmark for the industry's commitment to security, making the quality of AI safeguards transparent to end-users and policymakers.

In conclusion, as AI technologies advance, the disparities in security measures revealed by FAR.AI's research highlight an essential area for further exploration, advocacy, and reform in AI safety protocols. Harnessing these insights can lead to a more secure future where AI systems serve beneficial societal roles without posing risks to public safety.