Mythos bolsters case for coding's LLM fit

Hot take: this is a real signal, but it is not proof that coding is uniquely vulnerable; it is proof that coding is one of the clearest places where LLMs get tight feedback, abundant training data, and measurable wins.

• –Mythos’ security results matter because they target a hard, high-value task with clear success criteria: find a bug, prove it, exploit it, and verify it.
• –Coding is unusually LLM-friendly because it is text-heavy, highly structured, richly documented, and full of local patterns that models can interpolate well.
• –The post is strongest when it contrasts coding with domains where success is fuzzier, data is scarcer, or embodiment and long-horizon judgment matter more.
• –The leap from “great at finding bugs” to “SWE is disproportionately at risk” is plausible, but it overgeneralizes from a narrow slice of software work to the whole profession.
• –The likely near-term displacement is in repetitive repo-local work: refactors, test generation, static analysis, triage, and vulnerability hunting; architecture, product judgment, and cross-team coordination are harder to automate.