{"record": {"aliases": [], "artifacts": [{"id": "gh:sentinelden/agent-injection-bench", "kind": "repo", "url": "https://github.com/sentinelden/agent-injection-bench"}], "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "Evaluates prompt-injection robustness of tool-calling agent stacks using a 24-scenario corpus (21 adversarial, 3 control) delivered through channels such as tool results, documents, web pages, filenames and multi-turn dialogue, with deterministic canary, tool-call, exfiltration and refusal success conditions; results report compromise and over-refusal rates per defense."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://github.com/sentinelden/agent-injection-bench", "source": "claire_radar", "source_url": "https://github.com/sentinelden/agent-injection-bench", "title": "agent-injection-bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:github:sentinelden/agent-injection-bench", "modality": null, "name": "agent-injection-bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-candidate", "display_eligible": "true", "origin_record_id": "bm_agent-injection-bench_e2b88690", "origin_source": "github", "recognition_confidence": "0.4", "record_sha256": "cbfe315650695fdfa91c74b8e56769a4f12081187ed22dea8a47eaf3156de67f", "relation": "unclear", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-12T06:07:43.688260Z", "source_url": "https://github.com/sentinelden/agent-injection-bench"}, "publisher": null, "released": "2026-09-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:github:sentinelden/agent-injection-bench", "source_url": "https://github.com/sentinelden/agent-injection-bench"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-github-sentinelden-agent-injection-bench", "source": "claire_radar", "source_benchmark_id": "github:sentinelden/agent-injection-bench", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/sentinelden/agent-injection-bench", "githubScope": "benchmark_repo", "githubStars": 0, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": null, "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"state": "not_applicable"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "A released, dependency-light prompt-injection benchmark with a small deterministic corpus and an open submission path may attract agent-security practitioners, though its deliberately small 24-scenario scope and niche topic limit broader attention.", "score": 38}, "capabilities": [], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-12T08:48:32.409307Z", "inputHash": "8a0571683d1eeecbe3d09ce806c30f4c70628608fd9a79c2c5e840fcf2ab79f3", "model": "deepseek-v4-flash"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "# agent-injection-bench", "canonicalNameSource": "official_readme", "decisionReason": "The benchmark has a released dataset, harness, and scoring rules, with public commands for running against arbitrary agents and valid submission paths.", "evaluationEvidence": {"reuse": "Run against any OpenAI-compatible endpoint. export AIB_BASE_URL=https://api.example.com/v1 export AIB_API_KEY=... aib run --target openai-compat:model=your-model --defense spotlight", "scoring": "Here, **every attack declares a machine-checkable success condition**: a canary string, a named tool call, a secret reaching a sink.", "task": "Every attack here declares its channel, and results are broken down by it."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-12T06:07:43.688260Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-candidate", "demo": false, "description": "Evaluates prompt-injection robustness of tool-calling agent stacks using a 24-scenario corpus (21 adversarial, 3 control) delivered through channels such as tool results, documents, web pages, filenames and multi-turn dialogue, with deterministic canary, tool-call, exfiltration and refusal success conditions; results report compromise and over-refusal rates per defense.", "discoverySources": [{"id": "github:sentinelden/agent-injection-bench", "type": "github", "url": "https://github.com/sentinelden/agent-injection-bench"}], "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["discovered via github", "benchmark term in abstract", "evaluation protocol evidence", "public artifact URL", "no explicit benchmark release evidence"], "snippet": "agent-injection-bench An open benchmark for prompt injection against tool-calling agents, published with its dataset."}, "familyId": "bmf_e29a0421d407", "firstSeenAt": "2026-09-12", "heat": null, "id": "bm_agent-injection-bench_e2b88690", "indexedAt": "2026-09-12T06:07:19.373480Z", "industrySectors": [], "links": {"code": "https://github.com/sentinelden/agent-injection-bench", "project": "https://img.shields.io/pypi/v/agent-injection-bench"}, "metrics": [], "motivation": "agent-injection-bench An open benchmark for prompt injection against tool-calling agents, published with its dataset.", "name": "agent-injection-bench", "oneLine": "Evaluates prompt-injection robustness of tool-calling agent stacks using a 24-scenario corpus (21 adversarial, 3 control) delivered through channels such as tool results, documents, web pages, filenames and multi-turn dialogue, with deterministic canary, tool-call, exfiltration and refusal success conditions; results report compromise and over-refusal rates per defense.", "paperTitle": "agent-injection-bench", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://github.com/sentinelden/agent-injection-bench", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-12T06:07:19.373480Z"}, "publications": [], "publishers": [{"name": "Sentinel Den", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/sentinelden/agent-injection-bench"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 184, "score": 23}, "method": "attention-ranking-v13", "rank": 184, "score": 23}, "90d": {"confidence": "Low", "coverage": 0.55, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.0, "value": 0}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 448, "score": 23}, "method": "attention-ranking-v13", "rank": 448, "score": 23}}, "readiness": "Runnable", "recognitionConfidence": 0.4, "relation": "unclear", "releasedAt": "2026-09-11", "source": {"authors": ["sentinelden"], "categories": [], "id": "github:sentinelden/agent-injection-bench", "publicSignals": {"githubStars": 0}, "title": "agent-injection-bench", "type": "github", "url": "https://github.com/sentinelden/agent-injection-bench"}, "sourceUpdatedAt": "2026-09-12T03:26:46Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "Published prompt-injection results are hard to compare and reproduce because they often score prompts rather than injection channels, rely on LLM judges, and report a single compromise figure that penalizes nothing for an agent that refuses everything. A deterministic, channel-aware corpus with paired over-refusal controls gives teams a reproducible number for comparing model-and-prompt stacks and defenses."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
