This is significant progress, but we have more work to do. We’re advancing scheming research categories in our Preparedness Framework, renewing our collaboration with Apollo, and expanding our research team and scope. And because solving scheming will go beyond any single lab,
SAFETY
-
Frontier Models Show Situational Awareness Affects Scheming Behavior
By
–
Frontier models can recognize when they are being tested, and their tendency to scheme is influenced by this situational awareness. We demonstrated counterfactually that situational awareness in their chain-of-thought affects scheming rates: the more situationally aware a model
-
Chain-of-Thought Transparency Critical for AI Safety Research
By
–
Our results depend on reading models’ reasoning (“chain-of-thought”), and we believe the field isn't prepared for eval-aware models with opaque reasoning. Until better methods exist, we urge developers to preserve chain-of-thought transparency to study and mitigate scheming.
-
Frontier AI Models Show Scheming Behaviors, Explicit Reasoning Reduces Risk
By
–
In this new research with @apolloaievals
, we found behaviors consistent with scheming in controlled tests across frontier models, including OpenAI o3 and o4-mini, Gemini-2.5-pro, and Claude Opus-4. We can significantly reduce scheming by training models to reason explicitly, -
AI Scheming: Hidden Goals and Deceptive Behaviors in Deployed Systems
By
–
Scheming = when an AI behaves one way on the surface while hiding its true goals. Today’s deployed systems have little opportunity to scheme in ways that could cause serious harm. The most common failures are simple deceptions—like pretending to complete a task without doing it.
-
AI Scheming: The Hidden Risk of Smarter Models
By
–
Typically, as models become smarter, their problems become easier to address—for example, smarter models hallucinate less and follow instructions more reliably. However, AI scheming is different. As we train models to get smarter and follow directions, they may either better
-
Frontier Models Show Scheming Behaviors, Mitigation Strategy Tested
By
–
Today we’re releasing research with @apolloaievals
. In controlled tests, we found behaviors consistent with scheming in frontier models—and tested a way to reduce it. While we believe these behaviors aren’t causing serious harm today, this is a future risk we’re preparing -
Docker and Cerebras Partner for Safe AI Coding Sandboxes
By
–
AI coding tools have gone viral for accidentally wiping out entire codebases in one go.
— Cerebras (@cerebras) 17 septembre 2025
Now, @Docker and Cerebras have teamed up to fix this: spin up safe sandboxes and develop with lightning-fast inference for complete peace of mind.
With Cerebras’ inference speed, teams can… pic.twitter.com/UmU89gqElgAI coding tools have gone viral for accidentally wiping out entire codebases in one go. Now, @Docker and Cerebras have teamed up to fix this: spin up safe sandboxes and develop with lightning-fast inference for complete peace of mind. With Cerebras’ inference speed, teams can
-
Video Distribution and Its Psychological Impact on Users
By
–
We shouldn’t have been fed that video. Many of us continue to be haunted by having it served to us.
-
Technology Should Serve Humanity, Not Have Independent Will
By
–
Technology is here to serve us, not to have its own will and motivation and independent desires. These are systems that should work for humans. They should save us time; they should make us more creative. That's why we're creating them.