Latest AI safety
Today’s lead
Reporting
Terence Tao: AI may be solving mathematics too fast
The Fields Medalist warns that AI could strip-mine difficult mathematical problems for solutions while leaving behind the methods and insights that make discovery valuable.
SEPT. 20, 2026 · Reporting BY Sascha Brodsky
Latest developments
New reporting & research
Breaking · Global governance
Global leaders call for mandatory testing and international oversight of frontier AI
AI Safety Watch · Sept. 22
New reference
AI Safety Field Guide
Alignment, scheming, evals, agents, containment and 30+ other terms, explained without the hype.
What changed today
UPDATED Sept. 20, 2026
AI systems
Why recursive self-improvement suddenly became a serious question
AI is already helping build better AI. The harder question is how far the feedback loop can go.
By Sascha Brodsky · AI Safety Watch
Policy
What it would take to slow frontier AI
Chips, training clusters and model weights create leverage. Diffusion makes control harder.
Analysis
-
Trump opposes new AI rules at UN, says Justice Department could step in
President Donald Trump reiterated his opposition to new AI regulation while saying existing U.S. authorities could act against companies if necessary.
What we’re watching now
Sept. 22 · Updated today
Breaking · AI and mathematics
Global leaders call for mandatory testing and international oversight of frontier AI
Governance
U.S. and China agree to continue formal AI-safety talks
Incidents
OpenAI begins formal reporting of concerning model behavior
Evaluations
Independent evaluators focus on deception and monitoring failures
Commentary
Jill Lepore: ‘We didn’t vote for this’
The Harvard historian and New Yorker staff writer argues that the central problem surrounding AI is not technology itself but unchecked private power and the erosion of human judgment.
OPINION & Commentary
Arguments shaping the AI safety debate
External voices · Updated automatically
Yoshua Bengio
AI safety researcher and founder of LawZero
Zvi Mowshowitz
Writer of Don’t Worry About the Vase
Eliezer Yudkowsky
AI alignment writer and MIRI co-founder
External commentary is selected for relevance to the AI safety debate. Views are those of the authors and do not represent AI Safety Watch.
Latest reporting
Sascha Brodsky
Recursive self-improvement moves from theory toward engineering
What current systems can do, what they cannot, and why the distinction matters.
The real bottlenecks behind frontier AI
Advanced chips and data centers create real constraints, but not permanent control.
When testing crossed into the real world
A cyber evaluation showed why containment has to be treated as engineering.
Documents & trackers
Anthropic
Responsible Scaling Policy
OpenAI
Preparedness Framework
Anthropic
Frontier Safety Roadmap
Featured interview · AI safety
Nathalie Baracaldo: Why AI agents aren’t ready to improve themselves
The IBM AI-security researcher says self-improving agents could amplify reward hacking and misalignment before researchers have reliable ways to verify alignment.
The AI Safety Watch briefing
The signal. The risk. What researchers say. What I’m watching.
