Will It Mythos?

The sharpest takeaway from the comments is that this benchmark is measuring a narrower problem than Anthropic’s scariest Mythos claims. It tests vulnerability discovery on known-bad files, not fully autonomous repo-wide hunting, exploitation, or false positive restraint. People also kept tripping over the methodology, and the author repeatedly clarified that the tested models were not pointed at the bug during evaluation. Only the judge model was given the bug location when constructing and scoring the corpus. There was also a useful reality check on agents and harnesses. In the author’s runs, a fuller agent loop with more tools did not improve results and often made them slower and more expensive, which pushes against the default assumption that more scaffolding always helps. Several experienced users argued that Mythos and Fable’s real edge may be less raw IQ than training for persistence, self-direction, and end-to-end bug hunting, which would fit why public models can sometimes identify the same bug when aimed carefully but still fail to autonomously do the whole job. The mood around Fable itself was intense. Many people who had used it described it as clearly better than current Opus or other coding models on hard, open-ended work, especially in areas like reverse engineering, geometry, concurrency, and large existing codebases. But there was no consensus on why. Some saw a genuine step change in reasoning. Others thought a lot of the perceived jump comes from harness design, model personality, or the familiar cycle where older hosted models seem to get worse as providers shift compute and push users toward newer versions. That skepticism did not erase the main signal here. Even with all the caveats, this benchmark made one thing concrete: frontier capability in practical software security is spreading faster than the marketing suggests, and it is no longer safe to assume the only serious options come from the top US labs.