Claude Mythos breaks METR benchmark ceiling
claude mythos preview hit the ceiling of metr's test methodology with a 50 percent success rate on 16-hour tasks.
metr reported that claude mythos preview reached the upper limit of its current testing suite, estimating a 50 percent task completion probability on work taking 16 hours. the organization noted that measurements at this level are unstable due to sparse task coverage, necessitating new benchmarks for more capable models. palo alto networks validated these capabilities, revealing that models completed a year's worth of manual penetration testing in just three weeks. the firm warned that frontier models act as autonomous agents capable of chaining low-rated vulnerabilities into critical attack paths within 25 minutes.