Palo Alto Networks found 75 vulnerabilities in its products within a month of starting to use high-performance AI cybersecurity models from Anthropic and OpenAI, Axios reported on Tuesday. That was more than seven times its average monthly discovery rate of 5 to 10.

Palo Alto Networks is one of the few companies that can access both Anthropic's Mythos preview and OpenAI's GPT-5.5-Cyber.

Palo Alto Networks said it scanned more than 130 products over the past month and patched all 75 vulnerabilities it found. It added there were no cases of the vulnerabilities being exploited in real attacks.

Lee Klarich (리 클라리크), Palo Alto Networks' chief product officer, said the two models were particularly strong at linking multiple vulnerabilities to create a working attack path. He said there were several cases in which flaws not serious enough to disclose individually became high-risk vulnerabilities when combined. In internal tests, the models generated working exploits with a probability of more than 70 percent, he said. They write exploits far better than what he had seen before, he said.

But finding vulnerabilities using AI models still required human expertise and customisation. The average false-positive rate was about 30 percent, and results varied significantly depending on how researchers trained the models and what contextual information they provided, Axios reported. Klarich said the models were not magic. He said the company invested significant time in building an "AI scanning harness" to connect the models to scan targets.