Gray Swan relaunches Arena with USD $300,000 challenge

Gray Swan has launched a redesigned version of its Arena AI red-teaming platform and unveiled a USD $300,000 challenge tied to the update.

Arena is intended to widen access to adversarial testing of advanced AI models. Gray Swan says the platform has a community of more than 15,000 participants worldwide.

The Pittsburgh-based company says the revised Arena uses a more game-like design to make AI red teaming easier for newcomers and experienced security researchers alike. The service now includes beginner-friendly and ongoing challenges, training certificates, and a route for top performers to join its wider red-teaming community for commercial testing work.

Gray Swan works with AI developers and public bodies including Anthropic, Google DeepMind, OpenAI, Meta, Amazon, the UK AI Security Institute, and the National Institute of Standards and Technology's Centre for AI Standards and Innovation to test models anonymously.

Broader access

Gray Swan argues that expanding participation in AI red teaming could help uncover weaknesses that conventional security testing might miss. Arena is aimed not only at established security specialists, but also at people from creative and non-traditional backgrounds who may approach model testing differently.

"Understanding the safety and security of AI systems is hugely important today, and it shouldn't be left to a small circle of specialists. Some of the most unexpected vulnerabilities are uncovered by people who don't fit the traditional security researcher mold. The Arena levels the playing field, opening the door for anyone to learn and compete while uncovering valuable insights into where AI systems fail so that we can all work towards deploying AI safely and securely," said Matt Fredrikson, Chief Executive Officer of Gray Swan.

The newly announced contest, Indirect Prompt Injection Black, is the largest Arena competition the company has run so far. It will offer USD $300,000 in prizes, with sponsorship from Anthropic, Meta, Google DeepMind, and Reflection AI.

The challenge focuses on indirect prompt injection, a type of attack in which an AI model is influenced by instructions embedded in external content rather than by a direct user prompt. Participants will be asked to find vulnerabilities across a series of scenarios to be released in stages.

Research pipeline

Arena is not only a competition venue but also a source of data for Gray Swan's research. Information gathered through the platform is used to build benchmarks and evaluations intended to help researchers study new AI risks and measure the safety and security of newly released models.

That research has been cited in system cards and model launch materials published by major AI labs, according to Gray Swan. The link between competition findings and formal research is central to the company's pitch to both model developers and corporate customers concerned about AI security.

Gray Swan also tied the Arena relaunch to its wider product set. Techniques identified through Arena feed into Shade, its testing platform for AI vulnerabilities, and Cygnal, its runtime defence product designed to detect and block attacks while enforcing operating policies.

Gray Swan describes those tools, together with Arena, as a continuous loop linking vulnerability discovery, system testing, and live protection. The model is intended to cover both proprietary frontier systems and open-weight models used by organisations.

Growing market

The relaunch comes as AI developers and regulators pay greater attention to model safety, misuse, and resilience against manipulation. Prompt injection and related attacks have become a growing concern as companies deploy generative AI systems in customer service, software tools, and internal business processes.

By combining public competitions with commercial testing and monitoring products, Gray Swan is positioning itself in a part of the AI market that sits between cybersecurity, model evaluation, and governance. Its founders come from Carnegie Mellon University, and the company says its customer base includes leading AI labs and enterprises.

Arena participants can use the platform to practise testing models, compete for prizes, and build credentials that may help them move into AI security work. Top performers may also be invited to contribute to real-world testing engagements through Gray Swan's broader red-teaming network.

The Arena community includes hackers, researchers, and AI safety practitioners from around the world, with more than 15,000 members connected through the platform and its associated online channels, according to the company.