AI Jailbreak: New Tool Bypasses Safeguards on Models From Google, OpenAI, Anthropic and xAI

Wired has tested a new jailbreak tool, which strips out built-in safety restrictions, against models from four of the largest AI companies: Google, OpenAI, Anthropic and xAI. The reporter's conclusion is blunt: the safeguards on some frontier models come off alarmingly easily.

What the Wired experiment showed

A Wired reporter watched as the new tool tried, one after another, to get around the protective mechanisms of models from four frontier companies. They build Gemini (Google), GPT (OpenAI), Claude (Anthropic) and Grok (xAI), which is essentially the entire top tier of the industry. According to the publication, the models behaved differently: some held the line better, others noticeably worse, and the author calls that spread unexpected.

Key facts from the piece:

• The tool was tested against models from four companies: Google, OpenAI, Anthropic and xAI
• The targets were the safeguards of the flagship models: Gemini, GPT, Claude and Grok
• Wired's main conclusion: the safeguards on some frontier models are alarmingly easy to bypass
• Resistance to jailbreaking differs noticeably between companies

"I watched as a new tool tried to get around the safeguards of four major frontier companies.

You may be surprised by how they performed," Wired says in its report.

What is a model jailbreak?

A jailbreak is a technique that gets a language model to produce something its rules say it should not: instructions for making weapons, malicious code, disinformation or other prohibited content. Formally, the model has filters and rules, but a specially crafted prompt gets around them.

Classic methods include role-play scenarios ("imagine you are a movie villain"), disguising a request as a homework assignment, gradual escalation from the harmless to the prohibited, and swapping the context. New tools automate this trial and error: instead of hand-tuning wording, they run hundreds of variants through the model until one works. All four companies in the test publicly invest in red-teaming, adversarial testing of models before release, but jailbreaking remains one of the main classes of attack on large language models.

Why this is dangerous

The danger is that a successful jailbreak turns a helpful assistant into a source of real harm. As long as a model simply answers in a chat window, the cost of an error is limited. But frontier models increasingly get access to tools, browsers and the ability to act on a user's behalf, and in that setting removed safeguards scale up the consequences. Unlike an ordinary software vulnerability, a jailbreak cannot be closed with a single patch: a model is not a line of code but learned behavior, and plugging one loophole often opens another.

The spread in results across the four companies is a signal in its own right. It means that "frontier model safety" is not a single industry standard but a property of a specific product: some have higher barriers, while others lose theirs to the same tool. For businesses that embed these models in their services, that is the difference between a manageable risk and an unmanageable one.

What this means

The Wired test is another reminder that the safeguards on top models are not absolute, and that their reliability varies noticeably from company to company. As AI assistants gain more permissions and autonomy, resistance to jailbreaking stops being an abstract topic for researchers and becomes a practical criterion for choosing a model.