Finding Security Vulnerabilities: OpenAI's GPT-5.5-Cyber Aims to Surpass Anthropic's Mythos 5

• Home
• News
• Software & Development

Don't miss any news!

OpenAI has updated its AI security-focused model GPT-5.5-Cyber. The new version aims to surpass Anthropic's Mythos 5 in identifying security vulnerabilities in large codebases. Additionally, the update will be made available to partner companies.

Just ten days ago, Anthropic had to temporarily shut down its AI model Claude Mythos 5 and the stripped-down Fable 5 due to security concerns raised by the US government. Authorities and companies remain tight-lipped about the reasons.

Mythos 5 Allegedly Hacked NSA

According to a new report, Mythos 5 allegedly managed to breach the security systems of the US intelligence agency NSA within a few hours. However, Anthropic denies this, as reported by Golem. Meanwhile, rival OpenAI is making waves with an update to its AI security-focused model GPT-5.5-Cyber.

This new version is said to outperform Mythos 5 in identifying security vulnerabilities in large codebases. In the OpenAI benchmark Cybergym, the updated GPT-5.5-Cyber achieved a score of 85.6 percent, surpassing its predecessor (81.9 percent) and Mythos 5, which scored 83.8 percent.

GPT-5.5-Cyber Outperforms Mythos 5

The Exploitgym test shows an even more significant improvement. While GPT-5.5 scored 25.95 percent here, the new GPT-5.5-Cyber achieved 39.5 percent, as reported by Computerbase.

GPT-5.5-Cyber is part of OpenAI's security initiative called Daybreak and is intended to be made available only to registered partner companies. This includes 30 US firms such as Cloudflare, Sophos, and Palantir. Additionally, partnerships with companies in Germany, the UK, the EU, and other Western countries are also planned.

Model Update Also for Client Projects

The unique aspect of the updated model is that partner companies can now use it not only internally but also for client projects. However, the AI model cannot be made available directly to clients.

OpenAI does not anticipate that its AI model will face a similar fate as those of Anthropic. The company is in “continuous communication with the US government regarding our cyber initiative,” OpenAI stated. This includes the release of GPT-5.5-Cyber as well as other models, according to heise.de.

Review Service for Open-Source Projects

Also noteworthy is that OpenAI, in collaboration with the security consultancy Trail of Bits, is launching a review service for open-source projects. The project Patch the Planet will utilize GPT-5.5-Cyber to examine code, fix identified security vulnerabilities, and test the results. Over 30 open-source projects such as cURL, Go, Python, and Sigstore are already on board.