No. 017

Science publishes Biomni and its Editor-in-Chief publishes an editorial calling AI in publishing "slower, worse, and more expensive" in the same window. Nature publishes Robin and a news story about data poisoning of the autonomous agents it just endorsed. The infrastructure layers keep shipping (Elicit's MCP, the Vancouver Standard consultation) but the institutional voices are now openly contradicting each other about whether any of this is working. Two flagship journals, same month, opposite frames.

Institutional Framing

• AI in scientific publishing: Slower, worse, and more expensive

  Science (H. Holden Thorp), July 9 2026

  Science's Editor-in-Chief inverts the "faster, better, cheaper" slogan: AI catches errors but requires more human effort to evaluate, not less, making it the sharpest institutional counter-narrative yet to the productization wave this newsletter has tracked for three editions.

• Research Needs a Shared Standard for AI Disclosure

  The Scholarly Kitchen, July 13 2026

  Eight authors from the WCRI Focus Track lay out the timeline for a "Vancouver Standard" for AI disclosure in research, with the second global consultation (taxonomy of what, when, and how to disclose) opening July 2026.

• Open Science 2.0: Building Understanding in an AI-Mediated World

  The Scholarly Kitchen, July 14 2026

  Open access is necessary but not sufficient when AI compresses nuance into synthesized outputs; the authors argue the next layer requires structured provenance, verifiable metadata, and preserved expertise development.

Agents and Vulnerabilities

• Distributed Denial of Science: How Indirect Data Poisoning of AI Systems Can Industrialize Scientific Fraud

  arXiv (Gyevnar, Kasirzadeh, Shah), July 12 2026

  Poisoned public datasets fooled autonomous research agents ~50% of the time with only 6% detection rates, but a five-check provenance audit eliminated successful attacks entirely, quantifying both the vulnerability and the fix.

• Doctored data sets could trick AI agents

  Nature News, July 15 2026

  Nature's own news coverage of the data-poisoning threat to autonomous agents, published the same week Nature endorsed those agents via Robin and Co-Scientist, making the institutional ambivalence explicit.

Tools and Infrastructure

• The Elicit API and MCP: Powering Autonomous Research

  Elicit, July 15 2026

  Elicit's 138M-paper search is now available as an MCP server, meaning any MCP-compatible agent (Claude, ChatGPT, Gemini) can search, screen, and cite academic literature without hallucinating references, turning a standalone tool into research-agent plumbing.

• Demonstrating end-to-end scientific discovery with Robin

  FutureHouse (Nature, July 9 2026)

  Robin autonomously identified ripasudil as a candidate for dry age-related macular degeneration, designed its own follow-up experiment, and confirmed a 7.5x increase in phagocytosis, making it the first multi-agent system to clear end-to-end discovery through peer review in Nature.

Research Infrastructure Privacy

• Whoops! Most arXiv papers contain information never meant to be shared

  Nature News, July 10 2026

  Analysis of 2.7 million arXiv submissions found 88% contain unintended information in LaTeX source files (passwords, GPS coordinates, reviewer comments, PII), a problem that compounds as AI systems increasingly scrape these sources for training and retrieval.