THE
SIGNAL
A weekly report on the AI developments worth following, with enough context to understand what changed and why it matters.
OpenAI and METR disclose a multi-agent security incident at unprecedented resolution
The system around the model takes over
The same report,
easier to read.
The PDF remains the permanent edition. The website uses the same reporting, but adds a better reading layout, searchable issues, expandable paper notes, and a comfortable way to move through a long report.
Papers worth your attention.
Each note covers the paper’s claim, method, results, practical use, and limitations in plain English.
Prime Agent: A Self-Improving RLM Harness
Long-horizon performance depends on an external computational and stateful membrane that lets a model inspect, transform, verify, recover, and persist work beyond its active conversation.
Explore the paper note ↗Do User-Authored Permission Policies Improve Protection Against AI Agent Overreach?
Plain-language reusable policies reduce runtime prompts but do not automatically protect users better than reviewing each action; the human tendency to choose “ask” and then approve creates a gap between stated preference and actual commitment.
Explore the paper note ↗From Static to Dynamic: Benchmarking Real-World Code Review with MCR-Bench
Static one-shot review benchmarks overestimate real code-review ability because models must track defects as code and discussion evolve across rounds.
Explore the paper note ↗Browse earlier
weeks.
Search by company, model, method, policy, or theme. Every web issue links back to its original PDF.