Concern 985 · Working draft · editorial review in progress
Evaluations for dangerous capabilities
Dangerous-capability evaluations test whether an AI can perform tasks that might enable serious harm and whether safeguards limit those abilities.
Understanding the concern
A useful evaluation needs a clear threat, realistic conditions and a way to judge results. Access to tools and the amount of help provided can change performance. Failure on a test may mean the system lacks an ability, or that the test did not bring it out.
A driving test should assess more than parking in an empty lot. Similarly, AI safety tests need to reflect the conditions in which harm could occur. METR's 2026 pilot examined agents together with their organizational setting, including opportunities and controls. Evaluations provide evidence for decisions; they do not certify that every untested use is safe. Reporting limitations helps people interpret results honestly.
An example
A model passes a restricted test, but the deployment gives it additional tools that were never part of the assessment.
Illustrative unless identified as a documented case.
Words in plain English
- AI
- Software that learns patterns from examples or uses rules to perform tasks such as recognizing images or generating text. It can make mistakes and does not understand every situation as a person would.
- Agent
- An AI system that can choose and carry out several steps toward a goal, using the tools and access people give it.
Evidence and its limits
The pilot covers participating developers and a dated window, with access and disclosure limits. It is not a universal safety certification.
METR · May 2026 · Independent risk assessment
Frontier Risk Report (February to March 2026) ↗ (opens in a new tab)Pilot assessment of internal agents at four participating developers, with access and disclosure limits. Conclusions concern the assessment window.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
Does this evaluation cover the actual tools, access and conditions planned for deployment?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Humility
Humility means recognizing the limits of our knowledge and being willing to learn from others. Expertise can be valuable without answering every question.
State where evidence is weak, invite criticism, and listen to people with different experience. Leave room to revise a decision when new information appears.
Ask together: What might we be missing, and who could help us see it?
Stewardship
Stewardship means caring for resources and systems that others depend on, including people who will use them later. It looks beyond the benefit of a quick launch.
Plan for maintenance, energy use, data protection, and eventual shutdown. Ask who will look after the tool when the original team moves on and how future users can avoid inheriting hidden costs.
Ask together: What are we responsible for caring for over time?
Hope
Hope means seeing worthwhile possibilities and taking steps toward them, even when the outcome is uncertain. It does not require ignoring harm or promising that everything will work out.
Name the improvement you want in people's lives, then identify one realistic action and the conditions it needs. Keep room to revise the approach when evidence changes.
Ask together: What better outcome is worth working toward, and what can we do next?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.