CARI reading · 2 Oct 2026
46.8 / 100
Contested control
Above appetitescorecard base 44.5 + news adjustment +2.28
vs previous reading+0.2
vs 7 days earlier+1.6
vs 30 days earlier–
The full dashboard for this day is not in the public archive.
Summary
CARI stands at 46.8 (Contested control), +0.15 on the day: scorecard base 44.5 plus a news adjustment of +2.28. Today's news moved Control (+1.17) and Capability (+0.91). The largest driver was "AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks". The largest mitigation was "CARAT: Do Materials LLMs Reason or Recite?".
Bottom line. Control is the most advanced dimension at 48.7 (above appetite); watch for its next tripwire.
Six dimensions
| Dimension | Level | Base | News | Today | Zone |
|---|---|---|---|---|---|
| Capability | 45.0 | 43.0 | +2.02 | +0.91 | Above appetite |
| Diffusion | 48.2 | 45.9 | +2.29 | +0.33 | Above appetite |
| Accident | 45.0 | 42.0 | +3.00 | +0.81 | Above appetite |
| Control | 48.7 | 45.7 | +3.00 | +1.17 | Above appetite |
| Governance | 45.8 | 44.3 | +1.46 | −0.10 | Above appetite |
| Misuse | 48.6 | 47.2 | +1.38 | −0.21 | Above appetite |
Source: daily briefing email.
Change from the previous reading
2 Oct 2026 Previous reading (1 Oct 2026)
Stories that moved the index
- AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks+0.1 pts · arXiv cs.AI, Oct 1, 2026
- AI Agents are Vulnerable to Radicalization+0.1 pts · arXiv cs.AI, Oct 1, 2026
- “We’re not going to shoot ourselves in the foot” over hack fallout, says OpenAI’s chief research officer+0.1 pts · MIT Tech Review, Sep 30, 2026
Showing the largest movers of 11 stories in that day's briefing.