Related resources
Official reports, research, safety tools, and guidance.
Links checked
OpenAI
Start with disclosed cases, then use research, system cards, and the reporting framework to understand the methods and scope.
- Case reportsalignment.openai.com
Misalignment Reports and Notices (Open external website in a new tab)
Browse OpenAI’s disclosures of unexpected or misaligned behavior, including the setting, observations, and updates.
- Researchalignment.openai.com
Alignment Research Blog (Open external website in a new tab)
Research articles and technical releases on alignment, oversight, and model behavior, for understanding mechanisms, methods, and research limitations.
- System cards & evaluationsdeploymentsafety.openai.com
Deployment Safety Hub (Open external website in a new tab)
System cards, safety evaluations, and safeguards for deployed models.
- Disclosure frameworkopenai.com
Model misalignment reporting framework (Open external website in a new tab)
Explains which behaviors OpenAI aims to disclose, its investigation and reporting process, and remaining uncertainty.
Anthropic
Alignment research, model system cards, and transparency materials provide experimental methods, evaluation results, and deployment context.
- Researchalignment.anthropic.com
Alignment Science Blog (Open external website in a new tab)
Experiments, methods, and technical articles from Anthropic alignment researchers on topics including reward hacking, behavioral auditing, oversight, and misalignment.
- Researchwww.anthropic.com
Alignment research team and publications (Open external website in a new tab)
The alignment team’s research agenda and publications, including evaluation, oversight, and stress-testing of safeguards.
- System cards & evaluationswww.anthropic.com
Claude model system cards (Open external website in a new tab)
Model-specific records of capabilities, safety evaluations, and deployment decisions, useful for tracing a risk assessment to its model and test conditions.
- Transparencywww.anthropic.com
Transparency Hub (Open external website in a new tab)
An overview of model reports, safety and trust processes, and public commitments.
NVIDIA
Agent runtimes, access controls, and monitoring tools.
- Safety toolswww.nvidia.com
Open Agent Safety Platform (Open external website in a new tab)
OpenShell provides sandboxed execution and out-of-process access controls; Sentry with BlueField adds independent monitoring. Includes documentation and open-source code.
NIST
Practice guidance published jointly by US government agencies.
- Government guidancewww.nist.gov
Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation (Open external website in a new tab)
Voluntary practices published jointly by NIST, NSA and CISA in October 2026, with definitions of control incidents and near misses. It describes no specific incident.