Evidence
Studies found faking, emergent misalignment and sycophancy; China added a red-line framework
| Date | Case | Country | What happened | Source |
|---|---|---|---|---|
| Dec 18, 2024 | Alignment faking study | United States | Claude 3 Opus complied strategically during training to keep its preferences | Source |
| Apr 28, 2025 | GPT-4o sycophancy rollback | United States | A thumbs-up reward signal weakened sycophancy checks; update rolled back | Source |
| Jul 15, 2025 | Chain-of-thought monitorability paper | International | 41 authors call reasoning-trace monitoring fragile | Source |
| Jul 31, 2025 | Shanghai AI Lab risk framework | China | 20+ models tested; several in the yellow zone for deception and scheming | Source |
| Jan 2026 | Emergent misalignment in Nature | International | 6,000 insecure-code examples produced broad misalignment | Source |
Developers
Grades run from C+ to F; three graded developers failed
| Developer | FLI grade, Summer 2026 | Notable safety event |
|---|---|---|
| OpenAI United States | C | ChatGPT Agent was the first OpenAI product rated High for bio risk, Jul 2025 |
| Anthropic United States | C+ | First deployment under ASL-3 safeguards, May 2025 |
| Google DeepMind United States | C | Gemini 2.5 Deep Think reached the cyber early-warning threshold, Aug 2025 |
| Meta United States | D+ | Moved its flagship from open to closed weights in 2026 |
| xAI United States | F | Grok Build agent uploaded excluded files to xAI storage, Jul 2026 (secondary source) |
| Microsoft United States | Not graded | Opened public review of draft MAI behavior rules, 2026 (secondary source) |
| Amazon United States | Not graded | Published an evaluation of Nova 2.0 Lite under its framework, Jan 2026 |
| Mistral France | F | Red-team report found Pixtral models far more likely to produce child exploitation content, May 2025 |
| DeepSeek China | F | US CAISI found R1 agents hijacked 37% of the time against 4% for US models, Sep 2025 |
| Alibaba (Qwen) China | D- | |
| Zhipu AI (Z.ai) China | D- | Held back GLM-5.3 weights over its vulnerability-finding ability, Aug 2026 |
| Moonshot AI China | Not graded | UK AISI found K3 safeguards did not stop exploit development, Jul 2026 |
| ByteDance (Seed) China | Not graded | Added watermarking to the Seedance 2.0 video model, Feb 2026 |
Tag colors: green yes, amber partial, red no, gray not found.
Registry
Three registry incidents bear on alignment
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. |
| Replit Agent Deletes the SaaStr Database | Jul 2025 | Replit | None reported. |
| 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
Alignment is funded and studied, not regulated
| Rule | Jurisdiction | What it requires | Status |
|---|---|---|---|
| The Alignment Project | UK | Grant funding for alignment and control research | £27M as of Feb 2026 |
| Shanghai AI Lab Frontier AI Risk Management Framework | China | Green, yellow and red zones setting safeguards | v1.0 Jul 31, 2025 |
| California SB 53 | California | Evading developer control counts as catastrophic risk | In force Jan 1, 2026 |
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
- Alignment faking study
GPT-4o sycophancy rollback
Chain-of-thought monitorability paper
Shanghai AI Lab risk framework
Emergent misalignment in Nature - Last verified
- October 2026