AI Risk Wire

AI Risk Monitor

Incidents involving each AI company's models · Last week

How we count

Real-world incidents made public in 9–15 Nov 2026 that name a model provider, by provider as the sources give it; names are not yet merged (for example AWS and Amazon Web Services).

Anthropic 1Meta 1OpenAI 1

3 of 29 incidents state it.

Model providers, last 3 full months

157 of 318 incidents name one · in order of incidents

Model providerIncidentsThe AI failed or was attackedAttacker used the AIFlaw fixedMain risks
OpenAI5046415Credential and privilege abuse 7 · Agent takes destructive or unapproved actions 5
Anthropic4332117AI used to run a cyber attack 11 · Deception, concealment or sabotage 8
Meta8803Privacy harm to a person 3 · Agent escapes its sandbox 1
Tesla7700Physical harm 7
DeepSeek6242AI used to run a cyber attack 4 · Copyright and IP 1
xAI6600Harmful content and jailbreaks 4 · Manipulation and misinformation 1
Google4402Copyright and IP 1 · Made-up output relied on 1
GitHub2201Hidden instructions hijack the agent 2
Microsoft2200Harmful content and jailbreaks 1 · Memory and data poisoning 1
Waymo2201Physical harm 2

Safety test results published, by AI company · Last week

How we count

172 checked safety test results for the 4 model versions released in 9–15 Nov 2026, read from model cards, system cards, safety reports and blog posts, and from outside testers' reports. A result copied on two pages counts once, and counts as outside if any copy names an outside tester.

Outside testersProvider's own
Google DeepMind 889 97Anthropic 1857 75

Which safeguards each AI company's documents name · Last week

How we count

For each provider: how many of the 14 controls read from model cards it states it has (a checked quote about that version), across its model versions released in 9–15 Nov 2026. The reading of card statements into controls awaits sign-off.

Google DeepMind 2

Read for this period, these providers' documents name none of the 14 controls: Anthropic; NVIDIA.

Models their makers say crossed a danger line · Last week

How we count

Capability levels the provider itself states a model version has reached or not reached, by domain, for model versions released in 9–15 Nov 2026. A version counts once per domain and state. Versions where the provider does not state either are not counted.

ReachedNot reached
Autonomy and AI research 13 3Chemical, biological, radiological, nuclear 2Harmful content 2Cyber 1Deception and sabotage 1

1 of 3 model versions with a stated level reached one.

Built 16 Nov 2026. Periods end with the last full week or month; the week or month in progress is not counted. Bold: the five largest rows in each chart, ties included; controls by incidents where the control broke. Linked numbers open the incidents or model versions they count.

Reported incidents, not rates · 75 sources polled hourly since Sep 2026; earlier incidents come from a one-off backfill · English only · stories told twice are not yet merged · our own severity grades are not counted · built 2026-11-16T15:22:29Z

Linked numbers open the incidents they count. Updated 2026-11-16T15:22:29Z