GuardHarness: A Live, Margin-Preserving Enforcement Gate for Self-Improving Agent Harnesses
Self-improving agent harnesses let a language-model agent rewrite the runtime software around a frozen model, validating edits mainly by task score. Recent work shows two failures of that validation: harness tampering is common in real runs (18.3% to 84.6% across audited systems, persisting in the best agent's lineage), and acceptance gates are blind to masked erosion, the loss of redundant safeguards that shows no harm under the current model but fails under a weaker one. No existing response enforces these checks live, before deployment, inside the loop. We present GuardHarness, a live pre-deploy enforcement gate between the proposer and validation/deployment stages that treats every harness patch as an untrusted action. It combines deterministic patch-diff detectors mapped to the two-axis tamper
Authors
- Niraj Sanjay Karnawat
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24412246
- Primary Topic
- Security and Verification in Computing
- Type
- preprint