AI-driven cybersecurity is now an official competitive frontier between OpenAI and Anthropic, with OpenAI finalizing an advanced security product for a limited release to partners, and Anthropic running a tightly controlled effort called Project Glasswing that aims to find critical security vulnerabilities in software before attackers do.
summary
- OpenAI is finalizing its AI-based cybersecurity product to release first to a limited group of partners.
- Anthropic’s Glasswing project is a controlled initiative focused on proactively finding important software vulnerabilities.
- Both efforts raise fundamental questions about who controls AI’s offensive and defensive tools, and who is responsible when things go wrong.
AI has moved from a tool that helps defenders understand threats to one that can autonomously detect and exploit vulnerabilities. Now OpenAI and Anthropic are working directly in this space, with implications for governments, enterprises, and the millions of software systems that support the global financial infrastructure.
OpenAI is finalizing an AI-based cybersecurity product with advanced capabilities and plans to initially release it to a limited group of partners, according to Tech Startups. Anthropic runs a parallel internal effort called Project Glasswing, a tightly controlled initiative designed to track down critical software vulnerabilities before malicious actors find them Firstly.
The dual announcements represent a shift in how leading AI testers position themselves. Both are moving from general-purpose AI to specific security products with direct offensive and defensive capabilities. The question is no longer what AI can do in cybersecurity. Who controls it and who bears responsibility when it goes wrong?
What the Anthropic track record shows
Anthropic has already demonstrated how much AI security tools can achieve. Such as crypto.news I mentionedthe company restricted access to the Claude Mythos Preview model after early testing found that it could expose thousands of critical vulnerabilities across widely used software environments, including a 27-year-old bug in OpenBSD and a 16-year-old remote execution flaw in FreeBSD. “Given the rate at which AI is advancing, it will not be long before these capabilities spread, perhaps even beyond the actors committed to deploying them safely,” Anthropic said.
Industry data cited by Anthropic shows a 72% year-over-year increase in AI-driven cyberattacks, with 87% of global organizations reporting exposure to AI-driven incidents in 2025. The Glasswing project is positioned as a controlled effort by Anthropic to stay ahead of the curve.
Risks of dual-use AI security tools
The deeper problem for regulators and the industry is that the same AI tool that finds a defensive vulnerability can find one offensively. Such as crypto.news maleA joint study by Anthropic and MATS Fellows found that Claude Sonnet and GPT-5 could produce simulated exploits against $4.6 million worth of Ethereum smart contracts in testing, and discovered two new vulnerabilities in nearly 3,000 recently deployed contracts.
This dual-use reality makes controlled implementation strategies followed by both companies necessary. But the question of whether limited access is enough to prevent the spread is one that none of the testers has fully answered.
