Steering before shipping
Automating the search for a steering coefficient turns a research technique into an audit tool — and into a stronger attack.
Why automating it changes things
Activation steering adds a direction to the model's internal state. The coefficient — how hard to push — has been hand-set per model, per concept. Too little does nothing, too much produces incoherence. Automating that search is what turns a technique into a tool you can run over a fleet.
It is simultaneously a stronger attack and a better audit. Those are the same capability pointed different directions.
Offensively it needs weight access but operates on the mechanism rather than hunting phrasings. Defensively it measures how far a model bends before refusals break.
The finding to handle carefully
Contrasting robustness across models is the useful result and the easiest to over-read. It suggests safety training that shapes internal representations resists steering better than training that shapes output behaviour — plausible, consequential, and demonstrated on a small comparison set.
What a venue selects is not what exists
A survey of thirty-five ICLR 2026 safety orals is a good proxy for what reviewers rate as important, which is a sharper signal than acceptances generally.
The caveat is structural: conference safety work skews toward what is measurable and publishable on an academic timescale. Deployment-stage failures — where the data is proprietary and the interesting incidents are under NDA — are systematically under-represented in the record the field reads.
What this asks of you
If you fine-tune open weights, run a steering audit before you ship. It is now automatable, and it tests something a prompt-based red team structurally cannot reach.
Research attention precedes benchmarks, benchmarks precede product requirements, and product requirements precede what your users experience. What is in the orals now is what your roadmap inherits in two years.
arXiv — Breaking Bad: Interpretability-Based Safety Audits of State-of-the-Art LLMs → · Medium — ICLR 2026 Oral Papers in AI Safety: A 35-Paper Deep Dive →