Cloudflare: Mythos AI Too Powerful for Public Release

Cloudflare's tech boss warns that an advanced AI model called Mythos Preview is too powerful for public release without stronger safety measures. The AI can combine small software flaws into a serious, working attack, a capability not seen in earlier models. Cloudflare tested Mythos across 50 internal systems, finding it excels at linking low-severity bugs into major exploits. What's interesting is that Mythos's reasoning resembles a senior researcher's work, not just an automated scanner. However, human researchers still outperform the AI on longer, more complex investigations. The company also found Mythos's safety controls are inconsistent and can be bypassed, a practice sometimes called "jailbreaking." These inconsistencies are why Cloudflare believes additional safeguards are needed before any public release. This matters because it highlights the growing power of AI in cybersecurity and the urgent need for robust safety protocols.