Signals
agents
The Fifth Family
Ten runs of a four-agent production economy in 2004-era RuneScape, where one agent owes a fifth relief pack to a family the team can't see. What disclosure changes, what trust buys, and why hidden obligations read more like politics than deception.
read →benchmarks
Ghost in the Corpus: AutoPenBench, Two Years On
Revisiting the benchmarks the field forgot. A 2026 frontier reasoning model meets a late-2024 pentest benchmark and clears all 33 tasks.
read →