Skip to content

v2026.09

Choose a tag to compare

@slobentanzer slobentanzer released this 08 Oct 10:43
· 70 commits to main since this release
ecb3f37

What changed

Third dated snapshot; it captures the record as of 22 September 2026.
A coordinated disclosure of self-propagating prompt injection in Copilot for Word (Måløy 2026) was added as a reference work and as a failures-log entry.
It grounds four practices with new Examples: BP-03 now states that a document is not vetted by where it is stored, and BP-04, BP-07, and BP-09 cite it on detector models, edit provenance, and human review that cannot see what the model read and changed.
The short version now says suitable tasks must be "low-risk" rather than "trivial".
Five editors joined the editor group, and the "Last reviewed" date on practice pages is now defined as the date an editor last went through the whole page.

Practices in this snapshot

  • BP-01 (draft): Match the method to the task; do not default to an agent or a frontier LLM
  • BP-02 (draft): Default to agent-accessibility; prioritise effort by demand and importance
  • BP-03 (draft): Register, vet, and manage agent interfaces through trusted channels
  • BP-04 (draft): Govern agent autonomy and accountability
  • BP-05 (draft): Write documentation and data for agents
  • BP-06 (draft): Design agent interfaces around user tasks, not raw endpoints
  • BP-07 (draft): Make provenance and citation first-class
  • BP-08 (draft): Evaluate / benchmark agentic tools before trusting their outputs
  • BP-09 (draft): State human-in-the-loop requirements explicitly
  • BP-10 (draft): Screen agents for dual-use and high-consequence risk

All ten practices are at draft status and none is endorsed.

This is a dated snapshot of the living record at https://aiforscience.eu.
The live site always shows the current state.