Evidence
Three US labs flagged their own models at new risk levels in mid-2025
| Date | Case | Country | What happened | Source |
|---|---|---|---|---|
| May 22, 2025 | Anthropic activates ASL-3 | United States | First deployment under ASL-3 safeguards, for CBRN risk | Source |
| Jul 17, 2025 | ChatGPT Agent rated High for bio | United States | First OpenAI product given that rating | Source |
| Aug 1, 2025 | Gemini 2.5 Deep Think | United States | Reached the cyber early-warning threshold | Source |
| Sep 30, 2025 | CAISI evaluation of DeepSeek | United States and China | 19 benchmarks, three DeepSeek models against four US models | Source |
| Jun 30, 2026 | CAC filing totals | China | 988 services filed, 598 apps registered | Source |
Developers
Only four of 13 developers disclose pre-release testing by an outside evaluator
| Developer | Safety framework | Third-party testing | FLI grade, Summer 2026 |
|---|---|---|---|
| OpenAI United States | Yes Preparedness Framework v2, Apr 2025 | Yes UK AISI pre-deployment access | C |
| Anthropic United States | Yes Responsible Scaling Policy v3.4, Jul 2026 | Yes UK AISI cyber evaluation | C+ |
| Google DeepMind United States | Yes Frontier Safety Framework v3.1, Apr 2026 | Yes UK AISI, Apollo Research and others | C |
| Meta United States | Yes Superintelligence Scaling Framework v2.1, Oct 2026 | Yes Apollo Research | D+ |
| xAI United States | Yes Risk Management Framework, Aug 2025 | Unknown | F |
| Microsoft United States | Yes Frontier Governance Framework, Feb 2025 | Unknown | Not graded |
| Amazon United States | Yes Frontier Model Safety Framework, Feb 2025 | Partial independent evaluations published | Not graded |
| Mistral France | No | Unknown | F |
| DeepSeek China | No | No CAISI evaluated after release | F |
| Alibaba (Qwen) China | No | Unknown | D- |
| Zhipu AI (Z.ai) China | No | Unknown | D- |
| Moonshot AI China | No | No UK AISI assessed after release | Not graded |
| ByteDance (Seed) China | No | Unknown | Not graded |
Tag colors: green yes, amber partial, red no, gray not found.
Registry
Three registry incidents bear on model safety
Registry incidents where this discipline was tested in practice. Each links to its full record.
| Incident | Date | Developer | Regulator action |
|---|---|---|---|
| Claude Opus 4 Blackmails an Engineer in a Test | May 2025 | Anthropic | None. |
| Chinese State Group Runs Espionage Through Claude Code | Sep 2025 | Anthropic | On November 26, 2025, the House Homeland Security Committee asked Anthropic CEO Dario Amodei to testify. |
| OpenAI Test Models Breach Hugging Face | Jul 2026 | OpenAI | Senator Josh Hawley wrote to OpenAI on September 9, 2026. The White House Accord on Super Intelligence, signed September 29, 2026, commits signers to controls that keep models from hacking systems in unintended ways. No enforcement action was found. |
Rules
Lab frameworks set the thresholds; China requires filing before launch
| Rule | Jurisdiction | What it requires | Status |
|---|---|---|---|
| Anthropic Responsible Scaling Policy | Company | Capability thresholds trigger stronger safeguards | v3.4, Jul 2026 |
| OpenAI Preparedness Framework v2 | Company | High bio or chem rating requires safeguards | Apr 2025 |
| Google Frontier Safety Framework | Company | Critical capability levels with early warnings | v3.1, Apr 2026 |
| Interim Measures for Generative AI Services | China | Security assessment and filing before launch | Since Aug 2023 |
Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.
Record
Page Record
- Datasets
- ai_trust_evidence.csv (5 rows); registry incidents (3 linked); developer scorecard (13 developers, static table)
- Key sources
- Anthropic activates ASL-3
ChatGPT Agent rated High for bio
Gemini 2.5 Deep Think
CAISI evaluation of DeepSeek
CAC filing totals - Last verified
- October 2026