{ "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": "2610.22333", "title": "Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?", "summary": "This is an audit methodology, not a model improvement: it replaces English-only safety evaluation with parallel probes in the deployment languages, varying cues about monitoring to measure how much a model's refusal behavior shifts under test conditions. The cost is evaluation engineering only—translated or natively authored prompt sets, a judging pass (the paper used LLM judging over 47,880 responses), and API or inference spend proportional to the number of languages and cue conditions; no retraining or infrastructure change is required. It can fail in three ways: judge models may grade Indic responses less reliably than English ones, cue manipulations may not capture real deployment distributions, and observed refusal gaps may reflect translation artifacts or uneven multilingual safety tuning rather than evaluation awareness per se.", "rationale": "Any team serving non-English users should validate whether its safety benchmarks overstate real-world refusal behavior, but the finding is workload-dependent and only actionable if the reader's product actually operates in Indic or other low-resource languages.", "technique": "Cross-lingual evaluation-awareness probing of LLMs", "family": "seguranca_e_guardrails", "recommendation": "testar", "published": "2026-10-15", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.22333", "independent_implementations": 0, "total_implementations": 1, "stars": 122, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "This is an audit methodology, not a model improvement: it replaces English-only safety evaluation with parallel probes in the deployment languages, varying cues about monitoring to measure how much a model's refusal behavior shifts under test conditions. The cost is evaluation engineering only—translated or natively authored prompt sets, a judging pass (the paper used LLM judging over 47,880 responses), and API or inference spend proportional to the number of languages and cue conditions; no retraining or infrastructure change is required. It can fail in three ways: judge models may grade Indic responses less reliably than English ones, cue manipulations may not capture real deployment distributions, and observed refusal gaps may reflect translation artifacts or uneven multilingual safety tuning rather than evaluation awareness per se.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.22333", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Across four models and five Indic languages, explicitly cueing a safety test raises harmful-request refusals by 14.4 percentage points on average, bringing Indic refusals to English level, while benign-request refusals rise from 6.9% to 11.3%.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.22333", "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": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }