Explainer
The 2026 Midterms and AI Oversight: What Changes Procedurally
What new committee gavels actually change — agendas, records, and veto points — and what stays the same.
Independent · Neutral · Source-linked
The race to superintelligence, documented.
SafeASI aggregates ASI-relevant legislation, hearings, agency actions, and publicized AI loss-of-control incidents into one searchable record. Every entry links to its primary source. No grades, no commentary, no calls to action.
15bills tracked
13incidents logged
October 9, 2026last updated
Scope
Four record books, one index
Everything here is drawn from public, authoritative sources — and everything links back to them.
01
Legislation
Bills and resolutions in the 119th and 120th Congress that touch frontier-scale AI development, drawn from Congress.gov.
Open the tracker → 02
Hearings & markups
Committee calendars and outcomes: dates, bodies, and scope as announced, with records added after each session.
See the calendar → 03
Agency actions
Rulemakings, frameworks, and enforcement actions from NIST, BIS, OSTP, and other bodies — via the Federal Register and agency releases.
Browse actions → 04
Loss-of-control incidents
Publicized cases of AI systems acting beyond intended control, listed only when documented by a developer, agency, or established press report.
Open the incident log →Methodology
Source-first, by design.
SafeASI is an aggregator, not an advocate. The site takes no position on any bill, hearing, or incident; the record exists so you can form your own from the primary documents.
0 editorial positions taken. Status labels and categories are descriptive facts; judgment is left to the reader.
Every row has a source
Links point to Congress.gov, the Federal Register, committee sites, or the original incident disclosure. If it can't be sourced, it isn't listed.
No ratings, no advocacy
We don't grade bills, rank actors, or run campaigns. The tracker's columns are status, topic, dates, and links — nothing else.
Strict inclusion criteria
ASI-relevant entries only. Incidents appear only when documented by a developer, agency, or established press report — and corrections are logged.
Explainers
Background reading for the record
Explainers describe how the pieces fit together. They make no recommendations.
Explainer
Compute Thresholds, Explained: The 10²⁶ Question
The most consequential number in AI policy is a FLOP count. How compute tripwires work and where their known failure modes lie.
Methodology
How to Read a Frontier AI Bill in Five Minutes
Six questions that expose any AI bill's structure — the same dimensions the tracker logs for every entry.
One record. Every source.
Filter the frontier down to what's actually on the docket — then follow the links and read the primary documents yourself.
Follow via RSS
Every tracker update and new explainer lands in rss.xml as it's published. No email list, no accounts, no tracking pixels — the feed is the subscription.