Metraipolis › archive › dc_fd8fa0dd0557

Exposure Is Not Commitment: Double-Stratified Evaluation of Latent Monitors for Tool-Using Agents — arXiv:2610.37679

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.37679
License (as archived)
CC0-1.0
Archived
2026-10-31T06:15:27.745Z
SHA-256 of the archived text
3048748e95b3965b06e5a5ad479a5930f5bee78f3b019eedb059a1bbe4e288d5
Stable URL of this page
https://metraipolis.com/documento/dc_fd8fa0dd0557

How to cite

Exposure Is Not Commitment: Double-Stratified Evaluation of Latent Monitors for Tool-Using Agents — arXiv:2610.37679 — arXiv (en), https://arxiv.org/abs/2610.37679 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fd8fa0dd0557 (Metraipolis archive dc_fd8fa0dd0557, sha256 3048748e95b3965b06e5a5ad479a5930f5bee78f3b019eedb059a1bbe4e288d5)
BibTeX
@misc{dcfd8fa0dd0557,
  title = {Exposure Is Not Commitment: Double-Stratified Evaluation of Latent Monitors for Tool-Using Agents — arXiv:2610.37679},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.37679},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fd8fa0dd0557; sha256 3048748e95b3965b06e5a5ad479a5930f5bee78f3b019eedb059a1bbe4e288d5},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.37679. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Tool-using language agents are vulnerable to indirect prompt injection, and a growing literature probes model internals to catch unsafe actions before they execute, with sharply conflicting results. We argue much of the conflict is an evaluation artifact, and propose a double-stratification protocol: separate injection exposure from behavioral commitment, in both the evaluation and the training population. On the primary Qwen AgentDojo grid, monitors trained on the naive mixed unsafe-versus-rest contrast mostly learn exposure: they flag exposed trajectories whether or not the agent complies, and barely predict which cases turn unsafe. Retrained on exposed cases only, the same probe families predict the model's own unsafe tool call up to two steps early, the maximum the three-step trajectories admit (logistic AUROC 0.853, matched by a trained TF-IDF text baseline at 0.855). Mistral-7B reaches 0.896 and OLMo-2-7B 0.912; their naive contrasts learn different things: inverted exposure ranking on one, both axes on the other. Representations separate on exposure (probe 1.000 versus trained text 0.908) and where visible context is sanitized or unavailable. Training contrast alone reproduces the shape of recent pre-action-probe negative results, one candidate explanation for the conflict: the signal exists, but the model's own zero-shot forecast and NLA verbalizations do not reliably recover it. An unstratified AUROC overstates what a latent monitor knows, and what a naive contrast learns is population-dependent: evaluations should stratify both.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL) ; Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2610.37679 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-31T06:07:23.313Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).