Development

  • InfoQ — OpenAI classified GPT-6 Astra as the first model meeting its Preparedness Framework’s “Critical” cybersecurity threshold, citing demonstrated ability to discover and exploit previously unknown vulnerabilities.

AI

  • arXiv (cs.CL) — New method flags reward hacking in LLM evaluations by monitoring internal representations rather than relying on output behavior alone.
  • arXiv (cs.CL) — Proposes attention-dispersion patterns as a diagnostic signal for detecting hallucinated LLM outputs.
  • Hugging Face Papers — “Agora” uses git repositories as shared memory so multiple AI agents can collectively run automated research.
  • Hugging Face Papers — “HypoEvolve” applies genetic-algorithm-style evolution across multi-agent LLMs to generate and refine scientific hypotheses.

General

  • Keys Not Included (via Hacker News) — New York and Virginia have issued signed PDF417 barcodes on driver’s licenses for years without ever publishing the public key needed to verify them; the author reconstructed valid signing keys for both states from real license samples, showing the gap is a willingness problem, not an engineering one.