While there are concerns that the ability of advanced artificial intelligence models to find software vulnerabilities will lead to cybercriminals exploiting these bugs on a large scale, in reality everything is different. In particular, less than 0.5% of the vulnerabilities discovered by Anthropic’s AI in Project Glasswing were actually used in attacks.
Image source: anthropoic.com
To be more precise, the hackers only exploited one of the 225 vulnerabilities Officially registered Patrick Garrity, a researcher at VulnCheck, said the CVE number was linked to the Glasswing project through an open channel. Shortly after Anthropic announced the program in April, it began tracking vulnerabilities discovered and officially reported by the Glasswing project. The company subsequently said that releasing Mythos, an advanced artificial intelligence model, into the public domain would carry high risks because it has never been more effective at detecting vulnerabilities. Preview versions of this model are only available to trusted Anthropic partners.
Meanwhile, Patrick Garrity only added bugs that were officially registered as part of the Glasswing project to his registry and compared them to the vulnerability index actually used by attackers. As of Monday, September 21, of the 225 bugs, only one has been exploited – a critical SQL injection vulnerability (CVE-2026-26980) in the Ghost platform. “There is a big difference between discovering vulnerabilities and whether they will be useful to attackers and whether they will be exploited by attackers. The main takeaway from this data is that the vulnerabilities discovered and disclosed by Anthropic do not have much of an impact; they do not appear to have any different consequences than a random sample of any other vulnerabilities.”the researchers concluded.
Note that while modern AI models have learned to find vulnerabilities effectively, they are not yet very good at fixing them. Researchers at 1Password Off-by-1 Labs showed that only 26% of OpenAI’s GPT-5.5 and Anthropic Opus 4.8 models were able to fully fix the vulnerability; in 54% of cases, they either did not fix the problem, created a new vulnerability, or both. “Artificial intelligence has significantly lowered the barrier to vulnerability discovery, but the biggest gap occurs in the downstream steps: coordination, prioritization, problem resolution and patch deployment – which still requires significant human involvement, and Anthropic agrees. It seems they only realized this after the project was launched.”” concluded Patrick Garrity.
If you find an error, select it with your mouse and press CTRL+ENTER.
