AI Risk Index

• Anthropic warns small AI models could self-replicate through IoT devices Ars Tech · risk 41
• 'Welcome to the AGI era': OpenAI launches GPT-6 Astra VentureBeat · risk 39
• Apple warns rogue AI agents could ‘conceal’ and ‘manipulate,’ causing unintended consequences for its business - Fortune GN AI Risk · risk 37
• CrowdStrike's new kill switch can stop rogue AI agents in their tracks, CEO says - CNBC GN AI Risk · risk 37
• CrowdStrike Lets Security Teams Kill Rogue AI Agents - PYMNTS.com GN AI Risk · risk 37

What it measures. The risk that artificial intelligence causes serious harm or acts beyond human control — read from what the world's newsrooms are reporting right now.

Where the data comes from. Our own corpus of 40+ free public news feeds, including dedicated AI coverage, scored twice a day. Every headline gets a risk score: agent escapes, self-replication and AI-enabled attacks score highest; capability jumps and misuse score in the middle; regulation and policy score low.

Three components. The share of AI headlines that read as risk events (40%), the severity of the strongest headlines (40%), and public attention on AI-safety topics against three years of history (20%). Smoothed so one loud day does not swing the index.

Built to keep working. No single provider can switch this index off: when an enhancement source is unavailable, the built-in scoring carries on and the number still publishes. Same zones as every WTM index: Calm, Stable, Moderate, High Tension, Extreme Fear.

Data: public API · field ai_risk · CC BY 4.0 · Informational only, not advice.

A new kind of tension. The AI Risk Index tracks how much of the world’s coverage is about artificial intelligence going wrong — misuse, capability jumps, regulation fights and loss-of-control stories — and how much public attention those stories are pulling.

Why we score it. AI risk is now a geopolitical topic. Chip controls, deepfakes in elections and AI in weapons are tension stories, not tech stories. Nobody was measuring them on one scale, so we built it.

How it is built. Forty percent from our news corpus, forty from the top headlines of the cycle, twenty from public attention. The result is smoothed so one viral story cannot spike it, and a real trend cannot hide.

• Misuse incidents — deepfake fraud, AI-driven cyberattacks, scams at scale.
• Capability jumps — a new model that does something nobody expected.
• Regulation and chip controls — export bans and rulebooks that split the world into blocs.
• Attention spikes — when the public starts searching for AI safety, the score follows.

Today: 26 — Calm. AI news is routine. No safety incidents of note.

Direction: it held flat over the last 7 days and fell 4 points over 30 days. Our record since 17 September averages 30.0, so today sits below its own normal.

Top story in the cycle: “Anthropic warns small AI models could self-replicate through IoT devices”.

The scale. Under 40 is background noise. 40–54 is normal. 55–67 means a live incident or a heated regulation fight. Above 67 means AI risk is a top-tier world story this week.

November 2026 so far. Average 26.2 over 5 days. Peak 27 on 2 November, low 25 on 3 November. Read the November story →

Last month averaged 28.0, so this month is running lower by 1.8. Every month on record →

What counts as AI risk?

Anything where AI is the cause of harm or the subject of a fight: fraud, cyber, military use, misinformation, and stories about systems acting beyond their design. Product launches on their own do not score unless they trigger safety debate.

Is this about job losses?

Only when the story is about social stress — strikes, protests, laws. Ordinary automation news is not tension.

Why is it smoothed?

AI news comes in bursts. The smoothing (a rolling average that weights recent days half) keeps one viral story from swinging the score ten points and back in a day.

Does it feed the World Tension Index?

No. It stands alone on purpose. Mixing a new signal into a score with a five-month track record would break comparability. We may add it after a full year of data.