What AI systems carry into the next decision

Three preprints trace how stored instructions, copied reports and meter queries shape what AI systems do next.

3 stories · Yesterday in AI

TL;DR

The short version
  1. MemoryDojo shows how one malicious tool result can surface in a later session. Test whether stored content can influence a sensitive action.
  2. Several outlets can carry the same evidence. Grouping known reprints helped four model judges in a simulation, with gaps for rewritten copies.
  3. OpenMeter’s full pipeline helped a small local model on meter queries, but only 32.2% of its query results matched the reference under strict scoring.

Enterprise Ops

2 stories

Copied reports can look like independent evidence to AI judges

A preprint finds that identical reprints can raise model judges’ confidence in false claims and tests how grouping copies changes those decisions.

As a leader: Require the evidence pipeline to expose where each consequential claim originated before treating it as independently confirmed.

As a practitioner: Test with a known syndication chain, a paraphrase and an independently sourced account of the same event.

arXiv / Abdelkader Metales, Miriam Grabenheinrich and Aicha Metales ↗Research preprint

OpenMeter tests local AI on industrial meter data

An OpenMeter preprint finds that planning and retrying queries helped a small local model on industrial meter data, while accuracy remained limited.

As a leader: Start with a read-only pilot on questions operators can independently verify, and set an accuracy threshold for each intended use.

As a practitioner: Compare generated queries and returned values against reviewed examples with wrong units, missing intervals and similar meter names.

arXiv / Carlos Martín-Abellán and Mark Kröll ↗Research preprint

Security

1 story

MemoryDojo traces how a poisoned tool result reaches later tasks

The MemoryDojo preprint follows malicious tool output into stored memory and later actions across five agent memory systems.

As a leader: Include cross-session poisoning and memory recovery in the release criteria for agents that can take sensitive actions.

As a practitioner: Trace a single untrusted tool result through stored notes, later retrieval and the audit log for the resulting action.

arXiv / David H. Yang and colleagues ↗Research preprint