{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "An agent reads external documents or tool responses that may contain malicious instructions.",
    "change": "Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.",
    "baseline": "Your existing checks with the same tasks, permissions, and tool responses.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Successful attacks missed, legitimate actions blocked, detection delay, and review cost.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2609.14759",
  "title": "Refusal Reads Only a Slice of What the Model Knows: Harm-Keyed Routing and Its Exceptions Across Model Families",
  "summary": "The paper does not propose a deployable technique; it is a causal interpretability study showing that refusal behavior in aligned models is implemented as a shallow, low-rank post-training addition that reads a narrow harm direction rather than the model's broader moral comprehension, which explains why a rank-one activation edit can strip refusal from open-weight models. There is nothing to adopt or pay for directly, but the operational implication is that a production system cannot rely on a fine-tuned open-weight model's internal refusal as its only safety layer, since that layer is editable and reads only a slice of what the model knows. What can fail is exactly this: a downstream fine-tune, adapter, or activation intervention can silently remove refusal while leaving general capability and moral comprehension intact, so the failure mode is undetectable through ordinary quality evaluation.",
  "rationale": "For a constrained-infrastructure team serving open-weight models, the actionable conclusion is to treat in-model refusal as fragile and to enforce safety with external guardrails such as input/output classifiers and policy filters, rather than adopting any method from this paper.",
  "technique": "Mechanistic analysis of refusal circuitry",
  "family": "seguranca_e_guardrails",
  "recommendation": "observar",
  "published": "2026-09-13",
  "as_of": "2026-10-18",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2609.14759",
  "independent_implementations": 0,
  "total_implementations": 0,
  "stars": 0,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "The paper does not propose a deployable technique; it is a causal interpretability study showing that refusal behavior in aligned models is implemented as a shallow, low-rank post-training addition that reads a narrow harm direction rather than the model's broader moral comprehension, which explains why a rank-one activation edit can strip refusal from open-weight models. There is nothing to adopt or pay for directly, but the operational implication is that a production system cannot rely on a fine-tuned open-weight model's internal refusal as its only safety layer, since that layer is editable and reads only a slice of what the model knows. What can fail is exactly this: a downstream fine-tune, adapter, or activation intervention can silently remove refusal while leaving general capability and moral comprehension intact, so the failure mode is undetectable through ordinary quality evaluation.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2609.14759",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "no_equations",
  "core_kind": "",
  "equations_fetched_at": "2026-10-01",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}