Alignment & Safety · Interpretability
With Goodfire's Silico debugger, AI lie detectors nearing production, and new safety fellowships at Anthropic and OpenAI, the field is building real-world infrastructure while the crucial question of what these evals actually measure persists.
Alignment · Interpretability
After years as a niche research discipline, mechanistic interpretability is now spawning startups, fellowship programs, and off-the-shelf debugging tools, though the hardest problems remain unsolved.
Security · Alignment
As exploit windows shrink to hours, AI red teaming shifts from quarterly checkpoints to continuous automation, yet blind spots in the methodology remain that tools alone cannot fix.
Alignment · Reading Lists
New alignment reading lists and literature surveys are reshaping the field's self-definition, while battles over which papers make the cut expose deeper fractures in AI safety research.
Alignment & Safety · Investigation
An MIT audit of 72 AI agent frameworks reveals a stark absence of safety disclosures and kill switches, while Anthropic’s unreleased Mythos model deepens the chasm between benchmark performance and real-world trust.
AI · Safety
The 84-page card disclosure published with Mythos 5 is the most detailed pre-deployment evaluation on record. The strongest version of the safety claim is also the narrowest.
AI · Analysis
Three of the five most-cited frontier benchmarks have had their public splits leak into training corpora since January. The score on the leaked one is not the score on the held-out one.