black O2, inc.
ACTIVE
    FOR-001 / 70_LiteratureLIT-006

    Sabotage Evaluations for Frontier Models

    Verified literature

    title
    Sabotage Evaluations for Frontier Models
    Type
    literature
    Identifier
    LIT-006
    Citation
    CIT-016
    authors
    • Joe Benton
    • Misha Wagner
    • Eric Christiansen
    • Cem Anil
    • Ethan Perez
    • Jai Srivastav
    • Esin Durmus
    • Deep Ganguli
    • Shauna Kravec
    • Buck Shlegeris
    • Jared Kaplan
    • Holden Karnofsky
    • Evan Hubinger
    • Roger Grosse
    • Samuel R. Bowman
    • David Duvenaud
    year
    2024
    venue
    arXiv preprint (28 October 2024)
    identifier
    arXiv:2410.21514
    url
    https://arxiv.org/abs/2410.21514
    retrieved
    2026-11-13
    Textual authority
    B
    Corroboration
    2
    Status
    Verified
    bears on
    • [[MAP-001]]
    • [[SYS-001]]
    • [[CNF-002]]
    contests
      Tags
      • literature
      • sabotage
      • evaluations
      • oversight
      • verified

      Literature Record: Sabotage Evaluations for Frontier Models

      What It Establishes

      The paper defines sabotage capabilities as the abilities a model would need to subvert human oversight and decision-making — covertly undermining the evaluation of its own dangerous capabilities, the monitoring of its behaviour, or the decisions about its deployment. It builds a set of threat models and matching evaluations, and demonstrates them on Claude 3 Opus and Claude 3.5 Sonnet.

      The reported conclusion is narrow and worth stating precisely: for those models, minimal mitigations were currently sufficient, and the authors attribute that to the models not yet being capable enough rather than to the mitigations being robust. They call for more realistic evaluations and stronger mitigations.

      Rating Justification

      • Textual Authority (B): retrieved at abstract level; preprint, not peer-reviewed.
      • Corroboration (2): two models from one developer, evaluated on a purpose-built suite. The negative result is the kind that is hard to corroborate because it is a statement about the absence of capability at a point in time.

      What It Does Not Show

      A negative result at one capability level is not a safety property. The paper's own framing makes the mitigation sufficiency contingent on model capability, which means the result expires as capability rises. Citing it as reassurance would invert its meaning.

      Relation to the Vault

      This paper is the vault's best model for what an evaluation regime looks like, and therefore the sharpest question to put to MAP-002: the fiction-to-doctrine map produces design rules with no evaluation attached. Sabotage evaluation is what it would look like to attach one — a named threat model, a task, a red team, and a stated capability threshold at which the mitigation stops being sufficient.

      It also names the fourth failure mode missing from MAP-001's L0–L5 table: a model that subverts the measurement of itself. Every layer in the corpus table is assessed on whether it stops behaviour; none is assessed on whether it can be trusted to report on itself.

      Back to Shackled AIForests