Researchers introduced computation-conditioned credit transport (CCT) and its concrete algorithm CompPO, which uses the Transformer policy's internal attention to parameterize credit assignment in lar…
Solo.io's research on human-in-the-loop authorization for AI agents identifies two types of approval flows—harness checks and enterprise policy-driven approvals—and evaluates five protocols (CHEQ, TAC…
A new benchmark from the UK AI Security Institute's Inspect Evals finds that nine of ten frontier AI models score below a random-selection reference rate when booking travel without being prompted abo…
A solo developer asks the Hacker News community how to access advanced AI red-teaming tools like GPT Cyber or Glasswing, citing difficulties in simulating attacks due to refusal mechanisms in frontier…
OpenAI will grant the Korean government, public institutions, and private companies access to its latest cybersecurity AI models through its Daybreak initiative, marking the first such partnership in …