AI Risk Monitor
Incidents involving each AI company's models · Last 6 months
How we count
Real-world incidents made public in May 2026 – Oct 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).
175 of 410 incidents state it.
Model providers, last 3 full months
157 of 318 incidents name one · in order of incidents
| Model provider | Incidents | The AI failed or was attacked | Attacker used the AI | Flaw fixed | Main risks |
|---|---|---|---|---|---|
| OpenAI | 50 | 46 | 4 | 15 | Credential and privilege abuse 7 · Agent takes destructive or unapproved actions 5 |
| Anthropic | 43 | 32 | 11 | 7 | AI used to run a cyber attack 11 · Deception, concealment or sabotage 8 |
| Meta | 8 | 8 | 0 | 3 | Privacy harm to a person 3 · Agent escapes its sandbox 1 |
| Tesla | 7 | 7 | 0 | 0 | Physical harm 7 |
| DeepSeek | 6 | 2 | 4 | 2 | AI used to run a cyber attack 4 · Copyright and IP 1 |
| xAI | 6 | 6 | 0 | 0 | Harmful content and jailbreaks 4 · Manipulation and misinformation 1 |
| 4 | 4 | 0 | 2 | Copyright and IP 1 · Made-up output relied on 1 | |
| GitHub | 2 | 2 | 0 | 1 | Hidden instructions hijack the agent 2 |
| Microsoft | 2 | 2 | 0 | 0 | Harmful content and jailbreaks 1 · Memory and data poisoning 1 |
| Waymo | 2 | 2 | 0 | 1 | Physical harm 2 |
Safety test results published, by AI company · Last 6 months
How we count
2141 checked safety test results for the 76 model versions released in May 2026 – Oct 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.
Providers with fewer than 10 results are not charted: MiniMax (6), Alibaba Qwen (5), InclusionAI (Ant Group) (3), Upstage (1), Xiaomi MiMo (1), poolside (1).
Which safeguards each AI company's documents name · Last 6 months
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 May 2026 – Oct 2026. The reading of card statements into controls awaits sign-off.
Read for this period, these providers' documents name none of the 14 controls: Alibaba Qwen; DeepSeek; Dots Studio; Mistral AI; Moonshot AI (2 of 4 versions not fully read); StepFun; Tencent Hunyuan; Upstage; Xiaomi MiMo; Z.ai (Zhipu AI); poolside.
Models their makers say crossed a danger line · Last 6 months
How we count
Capability levels the provider itself states a model version has reached or not reached, by domain, for model versions released in May 2026 – Oct 2026. A version counts once per domain and state. Versions where the provider does not state either are not counted.