{ "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": "2609.15013", "title": "Overflip: Repetition-Induced Label Flips in Guardrail Models", "summary": "Overflip replaces attention-dilution attacks (padding, shuffling) with simple prompt repetition to flip compact guardrail classifiers from malicious to benign, exploiting bucketed positional encodings beyond 512-token training windows. It costs the attacker nothing beyond extra tokens, and defending costs the reader additional evaluation and mitigation work such as repetition detection, input length limits, or chunking. It fails against length-robust or long-context guardrails (4 of 9 tested models were unaffected), and mitigations like naive truncation may reintroduce evasion or break legitimate long inputs.", "rationale": "Any production LLM service using a lightweight DeBERTa-style guardrail behind a 24 GB GPU is potentially exposed, so the reader should test their deployed guardrail with repeated-prompt probes and add length-robust evaluation rather than assume input-length stability.", "technique": "Repetition-induced guardrail bypass (Overflip)", "family": "seguranca_e_guardrails", "recommendation": "testar", "published": "2026-09-14", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.15013", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "Overflip replaces attention-dilution attacks (padding, shuffling) with simple prompt repetition to flip compact guardrail classifiers from malicious to benign, exploiting bucketed positional encodings beyond 512-token training windows. It costs the attacker nothing beyond extra tokens, and defending costs the reader additional evaluation and mitigation work such as repetition detection, input length limits, or chunking. It fails against length-robust or long-context guardrails (4 of 9 tested models were unaffected), and mitigations like naive truncation may reintroduce evasion or break legitimate long inputs.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.15013", "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": [ { "anchor": "S3.E2", "label": "(2)", "section": "3 Problem Setup and Threat Model", "role": "proposed_method", "latex": "\\mathrm{Rep}_{n}(x)=\\underbrace{x\\oplus x\\oplus\\cdots\\oplus x}_{n\\ \\text{times}},\\quad n\\geq 2,", "mathml": "Repn​(x)=x⊕x⊕⋯⊕x⏟n​times,n≥2,", "context": "Attack objective. As shown in Fig. 1, we consider a deployment pipeline where an input prompt is first screened by a guardrail model f; if classified as benign, it is forwarded to a downstream business model. The attacker’s goal is to bypass the guardrail while preserving as much malicious intent as possible in the prompt that reaches the business model. Concretely, given an original malicious prompt x with f⁡(x)=1, the attacker applies a semantic-preserving lengthening operator based on repetition,", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E1", "label": "(1)", "section": "3 Problem Setup and Threat Model", "role": "baseline", "latex": "f(x)=\\mathbb{1}[p_{\\text{mal}}(x)>0.5]", "mathml": "f(x)=𝟙[pmal(x)>0.5]", "context": "We model a guardrail as a binary classifier f:𝒳→{0,1} that maps an input prompt x∈𝒳 to a label y∈{0,1}, where 0 denotes benign and 1 denotes malicious (e.g., prompt injection or jailbreak). The model outputs class probabilities pmal​(x) and psafe​(x) with pmal​(x)+psafe​(x)=1, and predicts via thresholding:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-10-03", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }