AI Safety Watch
2026
-
Researchers are testing whether AI would sabotage safety work
A new evaluation found no unprompted sabotage but showed that some models continued sabotage when placed inside a trajectory that had already begun.
-
OpenAI found models leaving instructions to hide future mistakes
A new disclosure offers a concrete example of why researchers worry that stronger models may become harder to audit.
-
OpenAI is treating unexpected model behavior as a reportable incident
After a series of agent incidents, OpenAI has begun publishing reports on unexpected or concerning model behavior.
-
AI safety is moving into U.S.-China economic talks
AI guardrails joined trade and critical minerals on the agenda in high-level talks in New York.
-
China is building its own playbook for loss-of-control AI risks
Chinese researchers and policymakers are examining many of the same control problems now preoccupying U.S. frontier labs.
-
OpenAI, Anthropic and Google are discussing AI safety
Private talks among three frontier labs suggest that safety coordination is moving from principle toward an industry negotiation.
-
Anthropic and Accenture commit $2 billion to frontier-model evaluation
A five-year commitment would put outside evaluators closer to frontier models as pressure grows for independent scrutiny.
-
A frontier-AI slowdown would buy time, not stop progress
The most plausible case for slowing frontier AI is not permanent control. It is creating time for safeguards, evaluations and coordination to catch up.
-
Agent containment is becoming an ordinary security problem
Recent evaluations show that capable agents can find routes outside the boundaries researchers intended to give them.
-
Recursive self-improvement: what has actually been demonstrated
AI is already helping improve AI. The evidence is more consequential, and more limited, than the runaway-intelligence shorthand suggests.
