Interpretability tools were used to find the cause, not describe it
When unexpected adversarial behaviour appeared, OpenAI used mechanistic interpretability to compare models trained with and without the suspect data and identified the source. That is interpretability doing diagnostic work rather than producing illustrations.
The most consequential interpretability result this month is not a new technique. It is a use: when unexpected adversarial behaviours emerged, OpenAI used in-house mechanistic interpretability tooling to compare models trained with and without particular data and successfully identified the source of the malicious behaviour.
That is a different activity from what interpretability has mostly produced in public. Feature visualisations, circuit diagrams and attribution maps are explanatory artefacts — they show that structure exists and can be described. Tracing a specific behavioural defect back to a specific cause in the training corpus is diagnosis, and diagnosis is what makes a field operational rather than descriptive.
Anthropic's line of work has been moving the same direction. Its Microscope tooling identified features corresponding to recognisable concepts in 2024, and by 2025 was revealing whole sequences of features and tracing complete paths from prompt to response. The progression from "this feature exists" to "this is the path the model took" is the progression from anatomy to physiology.
The honest caveat is that both results are reported by the labs that produced them, using tooling nobody outside can run. Interpretability's credibility problem is the same as safety evaluation's, and it is not solved by better techniques — it is solved by access. Cross-lab evaluation is one partial answer to a problem interpretability has too.
OpenAI — Findings from a pilot Anthropic–OpenAI alignment evaluation exercise → · Zylos Research — AI Safety, Alignment, and Interpretability in 2026 →