Palo Alto Networks warns Claude Mythos and LLMs power...

Claude Mythos Preview has slipped past the measuring stick that METR, the AI‑risk outfit, has relied on for years. While the model’s 50 percent success rate on tasks calibrated to a 16‑hour human effort sounds modest, it marks the first time the test framework hits its ceiling. METR’s own numbers—an estimated 50 percent time horizon of at least 16 hours, with a 95 percent confidence interval ranging from 8.5 to 55 hours—suggest the model can now outpace the benchmarks that once defined capability.

But the ripple isn’t limited to academic metrics. Palo Alto Networks, after running its own trials, says frontier models like Mythos are already acting as autonomous agents, hunting software flaws and stitching them together into full‑blown attack paths. The firm notes the models crammed a year’s worth of manual penetration testing into three weeks, a pace that reshapes how quickly threats can be generated. As AI moves from tool to independent actor, the security community faces a new set of challenges—ones that traditional testing regimes may no longer capture.