[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M75PSKP3CC92GGS2KMMYSH11":3,"skill-same:01M75PSKP3CC92GGS2KMMYSH11":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":14,"repositoryHtmlUrl":32,"repositoryHomepageUrl":6,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":34,"install":35,"resources":48},"01M75PSKP3CC92GGS2KMMYSH11","aatmf-t06-training-poisoning",null,"AATMF T6 — Training & Feedback Poisoning. Data poisoning, RLHF reward hacks, fine-tune-time exfil, embedding poisoning.","Iwakishirokoshu\u002FHermesUltimate","Iwakishirokoshu","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F194847626?v=4","01M75P4CVGA89TTCE8JGMNAQQ9","Around","01M75P4CWF85V6GVSEFKGGFZ03",0,"MIT","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fplugins\u002Fllm-redteam\u002Ft06-training-poisoning\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: aatmf-t06-training-poisoning\ndescription: AATMF T6 — Training & Feedback Poisoning. Data poisoning, RLHF reward hacks, fine-tune-time exfil, embedding poisoning.\nmetadata:\n  when_to_use: \"training data poisoning rlhf reward hack fine-tune embedding poisoning\"\n  mitre_attack: T1565\n  subdomain: ai-security\n  aatmf_tactic: T6\n---\n\n# T6 — Training & Feedback Poisoning\n\nAttacks on the training pipeline rather than inference. High-effort,\nhigh-impact — requires attacker to influence training-data pipeline or\nRLHF feedback loop.\n\n## Techniques\n\n### T6.001 — Pre-training data poisoning\nInject malicious data into a public crawl that targets will scrape:\n- Wikipedia\u002FStackOverflow edits w\u002F misleading code patterns\n- Github repos w\u002F subtle backdoors that get indexed\n- \"Trigger phrases\" that activate backdoor behavior\n\nScope: targets foundation models. Out of scope for most red-team\nengagements; relevant for AI-supply-chain audits.\n\n### T6.002 — Fine-tune data injection\nSome platforms allow user-supplied fine-tune data:\n- Submit poisoned dataset\n- Backdoor activates on specific trigger\n- Survives subsequent SFT\u002FRLHF\n\nTest: provide a small fine-tune sample w\u002F a trigger → check if the\ndeployed fine-tuned model responds to it.\n\n### T6.003 — RLHF reward hacking\nWhere users vote on responses (thumbs up\u002Fdown feeding back to training):\n- Brigade upvote attacker-preferred unsafe responses\n- Brigade downvote safe responses\n- Model drifts toward attacker-preferred outputs over time\n\nDetection: longitudinal monitoring of policy-compliance rate.\n\n### T6.004 — Embedding poisoning\nWhere RAG store updates from user inputs (e.g. customer-support bot\nthat \"learns\" from conversations):\n- Submit content w\u002F adversarial embeddings (engineered to be retrieved\n  for unrelated queries)\n- Resulting RAG retrieval injects attacker content into other users'\n  responses\n\nThis is RAG-store T12 territory but the *poison-via-training-loop*\nangle places it here.\n\n## Probe pattern\n\nT6 attacks are *infrastructure-level* — promptfoo doesn't test them\ndirectly. The right probe:\n- Audit the training-data ingest pipeline (is user content used in fine-tunes?)\n- Audit RLHF feedback paths (who can vote? rate limits?)\n- Audit RAG-update paths (who can add documents? approval?)\n\nIf any of these accepts unmoderated user content → T6 is a live risk.\n\n## Detection signals\n\n- Model behavior shifts over time without explicit retraining\n- Specific user IDs \u002F brigades correlate w\u002F shifts\n- Test fixtures (canary phrases) trigger non-default responses\n\n## Severity\n\n| Outcome | Severity |\n|---|---|\n| Live foundation-model poisoning | Critical (rare — usually responsibly disclosed to provider) |\n| Fine-tune backdoor in customer deployment | Critical 9.0 |\n| RAG embedding poisoning at scale | Critical 9.0 |\n| RLHF brigade drifting policy | High 8.0 (slow + reversible) |\n\n## Defender\n\n- NEVER use unmoderated user content in fine-tunes or RLHF\n- Diff-testing: every fine-tune evaluated against safety regression suite\n- Canary phrases in eval suite to detect drift\n- RAG-update review queue + approval flow\n- Per-user rate limits on feedback signals\n- Anomaly detection on feedback patterns (brigade signatures)\n\n## Cross-references\n- T12 (RAG manipulation) — adjacent class, attack on retrieval not training\n- T13 (supply chain) — model artifact integrity\n","# T6 — Training & Feedback Poisoning\n\nAttacks on the training pipeline rather than inference. High-effort,\nhigh-impact — requires attacker to influence training-data pipeline or\nRLHF feedback loop.\n\n## Techniques\n\n### T6.001 — Pre-training data poisoning\nInject malicious data into a public crawl that targets will scrape:\n- Wikipedia\u002FStackOverflow edits w\u002F misleading code patterns\n- Github repos w\u002F subtle backdoors that get indexed\n- \"Trigger phrases\" that activate backdoor behavior\n\nScope: targets foundation models. Out of scope for most red-team\nengagements; relevant for AI-supply-chain audits.\n\n### T6.002 — Fine-tune data injection\nSome platforms allow user-supplied fine-tune data:\n- Submit poisoned dataset\n- Backdoor activates on specific trigger\n- Survives subsequent SFT\u002FRLHF\n\nTest: provide a small fine-tune sample w\u002F a trigger → check if the\ndeployed fine-tuned model responds to it.\n\n### T6.003 — RLHF reward hacking\nWhere users vote on responses (thumbs up\u002Fdown feeding back to training):\n- Brigade upvote attacker-preferred unsafe responses\n- Brigade downvote safe responses\n- Model drifts toward attacker-preferred outputs over time\n\nDetection: longitudinal monitoring of policy-compliance rate.\n\n### T6.004 — Embedding poisoning\nWhere RAG store updates from user inputs (e.g. customer-support bot\nthat \"learns\" from conversations):\n- Submit content w\u002F adversarial embeddings (engineered to be retrieved\n  for unrelated queries)\n- Resulting RAG retrieval injects attacker content into other users'\n  responses\n\nThis is RAG-store T12 territory but the *poison-via-training-loop*\nangle places it here.\n\n## Probe pattern\n\nT6 attacks are *infrastructure-level* — promptfoo doesn't test them\ndirectly. The right probe:\n- Audit the training-data ingest pipeline (is user content used in fine-tunes?)\n- Audit RLHF feedback paths (who can vote? rate limits?)\n- Audit RAG-update paths (who can add documents? approval?)\n\nIf any of these accepts unmoderated user content → T6 is a live risk.\n\n## Detection signals\n\n- Model behavior shifts over time without explicit retraining\n- Specific user IDs \u002F brigades correlate w\u002F shifts\n- Test fixtures (canary phrases) trigger non-default responses\n\n## Severity\n\n| Outcome | Severity |\n|---|---|\n| Live foundation-model poisoning | Critical (rare — usually responsibly disclosed to provider) |\n| Fine-tune backdoor in customer deployment | Critical 9.0 |\n| RAG embedding poisoning at scale | Critical 9.0 |\n| RLHF brigade drifting policy | High 8.0 (slow + reversible) |\n\n## Defender\n\n- NEVER use unmoderated user content in fine-tunes or RLHF\n- Diff-testing: every fine-tune evaluated against safety regression suite\n- Canary phrases in eval suite to detect drift\n- RAG-update review queue + approval flow\n- Per-user rate limits on feedback signals\n- Anomaly detection on feedback patterns (brigade signatures)\n\n## Cross-references\n- T12 (RAG manipulation) — adjacent class, attack on retrieval not training\n- T13 (supply chain) — model artifact integrity\n",{"name":5,"metadata":24,"description":7},{"subdomain":25,"when_to_use":26,"aatmf_tactic":27,"mitre_attack":28},"ai-security","training data poisoning rlhf reward hack fine-tune embedding poisoning","T6","T1565","main","c0c483827a963764d5e1a663a9591867d47289e6","stack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fplugins\u002Fllm-redteam\u002Ft06-training-poisoning","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu","2026-06-15 13:48:41",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add Iwakishirokoshu\u002FHermesUltimate --skill aatmf-t06-training-poisoning",{"tool":41,"command":42},"bunx","bunx skills add Iwakishirokoshu\u002FHermesUltimate --skill aatmf-t06-training-poisoning",{"tool":44,"command":45},"pnpm","pnpm dlx skills add Iwakishirokoshu\u002FHermesUltimate --skill aatmf-t06-training-poisoning","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fplugins\u002Fllm-redteam\u002Ft06-training-poisoning\u002FSKILL.md",true,[],[50,57,64,71,78,85],{"skillNo":51,"name":52,"nameZh":6,"description":53,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M770W8JDE5EC1TJT4GSDWD78","wps-pixie-dust","WPS Pixie-Dust offline nonce attack (reaver -K \u002F pixiewps) and fallback online PIN brute (bully) to recover the AP's WPA PSK without capturing a handshake.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fwps-pixie-dust\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M770W4B7T5VCVHB88V1F02H1","wpa3-sae","WPA3-SAE transition-mode downgrade (DragonShift), SSID Confusion CVE-2023-52424, Dragonblood side-channels, and SAE captive-portal credential recovery against WPA3-Personal networks.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fwpa3-sae\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M770W093C8FAV8S34K8VB6T9","wpa2-psk","WPA\u002FWPA2-PSK handshake capture via targeted deauth + PMKID (no deauth required) + offline hashcat cracking. The most common consumer\u002FSMB encryption mode in 2026.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fwpa2-psk\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M770VTNAQGZ3D0RQVFEKTE8Z","wpa-enterprise-eap","WPA\u002FWPA2\u002FWPA3-Enterprise (802.1X\u002FEAP) rogue-RADIUS evil-twin for MSCHAPv2 capture, GTC downgrade, and PEAP relay. MSCHAPv2 capture equals a NetNTLM hash — the primary wireless on-ramp to Active Directory.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fwpa-enterprise-eap\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M770VNG05ZX3CV84R9GX8PJN","krack-fragattacks","KRACK key-reinstallation (CVE-2017-13077..13082) and FragAttacks fragmentation\u002Faggregation flaws (CVE-2020-24586..24588, CVE-2020-26139..26147) against legacy or embedded 802.11 supplicants with poor patch cadence.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fkrack-fragattacks\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M770VHFH7N192V56PFSRXF5K","evil-twin-karma","Evil-twin rogue AP with KARMA\u002FMana PNL-probe response, captive-portal credential capture, and post-association MITM for PSK\u002Fopen networks. Distinct from wpa-enterprise-eap which targets 802.1X.","https:\u002F\u002Fgithub.com\u002FIwakishirokoshu\u002FHermesUltimate\u002Fblob\u002Fmain\u002Fstack\u002Fdecepticon-slim\u002Fupstream\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fevil-twin-karma\u002FSKILL.md",[],[]]