TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

benchmarks evals

The Security Agent Failed Before the Tested Capability Appeared

Checkpoint instrumentation separates long-horizon failures that happen before and after an agent reaches the relevant state.

Published Updated Story ID: mp-2026-08-24-007
Read the complete editionStory JSON

Summary

Checkpoint instrumentation separates long-horizon failures that happen before and after an agent reaches the relevant state.

In one 92-seed study, protocol-disambiguation guidance raised state observation for Gemini 2.5 Flash from 65.5 to 95.4 percent, but repeating the design with Gemini 3.7 Flash produced the opposite effect. The shifting bottleneck shows why final success alone cannot reveal which capability failed or whether it was exercised at all.

Why it matters

Checkpoint instrumentation separates long-horizon failures that happen before and after an agent reaches the relevant state.

Limits and context

  • The shifting bottleneck shows why final success alone cannot reveal which capability failed or whether it was exercised at all.

Key claims

  1. Checkpoint instrumentation separates long-horizon failures that happen before and after an agent reaches the relevant state.

    Qualification: The shifting bottleneck shows why final success alone cannot reveal which capability failed or whether it was exercised at all.

    Evidence: source-2026-08-24-007

Sources

  1. arXiv preprint 2608.20563arXiv · primary research

Corrections

No corrections have been recorded for this story.