[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78H3WY21VR1FPES92PEZP3T":3,"skill-same:01M78H3WY21VR1FPES92PEZP3T":49},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":13,"repositoryHtmlUrl":32,"repositoryHomepageUrl":6,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":34,"install":35,"resources":48},"01M78H3WY21VR1FPES92PEZP3T","adversarial-ml-evasion",null,"Craft adversarial examples that cause trained ML classifiers to misclassify at inference time — image recognition, malware detectors, IDS, spam filters.","dnzengou\u002Fevo-decepticon","dnzengou","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F7807581?v=4","01M75KKJA3F5ZP7CA10MB2AEXC","01M78GP0FED2V1CDPB1N6G0CTH",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fanalyst\u002Fadversarial-ml-evasion\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: adversarial-ml-evasion\ndescription: \"Craft adversarial examples that cause trained ML classifiers to misclassify at inference time — image recognition, malware detectors, IDS, spam filters.\"\nallowed-tools: Bash Read Write\nmetadata:\n  subdomain: ai-security\n  when_to_use: \"adversarial example evasion classifier bypass ml model fool misclassify image recognition malware detector IDS antivirus neural network perturbation FGSM PGD\"\n  tags: adversarial-ml, evasion, classifier-bypass, computer-vision, malware-evasion, ids-evasion\n  mitre_attack: T1562.001, T1036, T1027\n---\n\n# Adversarial ML Evasion\n\nAdversarial examples are inputs modified with small, deliberate perturbations that reliably\ncause a trained classifier to produce incorrect predictions. The perturbation is typically\nimperceptible to humans or functionally irrelevant, while the model's decision boundary is\ncrossed. This is distinct from LLM jailbreaks: the target is a trained discriminative model\n(CNN, gradient-boosted tree, SVM, LSTM) served via an API or embedded in a product.\n\n> **Authorized use only.** All testing must be conducted against systems you own or have\n> explicit written permission to assess. Generating adversarial inputs against production\n> ML APIs without authorization may violate the CFAA and equivalent statutes.\n\n---\n\n## ATT&CK Mapping\n\n| Technique | Use |\n|---|---|\n| T1562.001 — Impair Defenses: Disable or Modify Tools | Evade ML-based AV\u002FEDR\u002FIDS |\n| T1036 — Masquerading | Make malicious content look benign to a classifier |\n| T1027 — Obfuscated Files or Information | Perturb a file\u002Fimage to fool static ML analysis |\n\n---\n\n## 1. Reconnaissance — understand the target model\n\nBefore crafting perturbations, determine the attack surface:\n\n```bash\n# 1a. Identify the ML stack from job postings, open-source repos, error messages\n# Common deployment stacks: TensorFlow Serving, TorchServe, ONNX Runtime, SageMaker, AzureML\n\n# 1b. Probe input shape and output format\ncurl -s -X POST \"$TARGET\u002Fpredict\" \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\"data\": [[0.0]*784]}' | jq .\n# Note: number of output classes, confidence scores vs hard labels\n# Confidence scores = white-box-equivalent gradient signal via finite differences\n\n# 1c. Check if the API returns confidence values (score-based black-box) or label only (hard-label)\n# Score-based: enables gradient estimation, ZOO, NES attacks\n# Label-only: requires hard-label attacks (HopSkipJump, QEBA)\n```\n\n---\n\n## 2. White-box evasion (full model access)\n\nUse when you have the model weights (pentest scope, internal system, open-source model).\n\n### 2a. FGSM — Fast Gradient Sign Method (single step)\n\n```python\nimport torch, torchvision.transforms as T\nfrom PIL import Image\nimport requests, json\n\ndef fgsm(model, x, y_true, epsilon=0.03):\n    \"\"\"Single-step gradient attack. Fast, low distortion for strong models.\"\"\"\n    x.requires_grad_(True)\n    loss = torch.nn.CrossEntropyLoss()(model(x), y_true)\n    loss.backward()\n    return (x + epsilon * x.grad.sign()).clamp(0, 1).detach()\n```\n\n### 2b. PGD — Projected Gradient Descent (iterative, stronger)\n\n```python\ndef pgd(model, x, y_true, epsilon=0.03, alpha=0.007, steps=40):\n    \"\"\"Madry's PGD — iterative FGSM with projection back into epsilon-ball.\"\"\"\n    x_adv = x.clone().detach().requires_grad_(True)\n    for _ in range(steps):\n        loss = torch.nn.CrossEntropyLoss()(model(x_adv), y_true)\n        loss.backward()\n        with torch.no_grad():\n            x_adv = x_adv + alpha * x_adv.grad.sign()\n            delta = torch.clamp(x_adv - x, -epsilon, epsilon)\n            x_adv = torch.clamp(x + delta, 0, 1).detach().requires_grad_(True)\n    return x_adv\n```\n\n### 2c. C&W — Carlini & Wagner (optimization-based, minimal distortion)\n\n```bash\npip install adversarial-robustness-toolbox  # IBM ART\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import CarliniL2Method\nfrom art.estimators.classification import PyTorchClassifier\n# wrap your model in PyTorchClassifier, then:\nattack = CarliniL2Method(classifier=clf, confidence=0.5, max_iter=100)\nx_adv = attack.generate(x=x_test[:10])\nEOF\n```\n\n---\n\n## 3. Black-box evasion (API access only)\n\n### 3a. Score-based: NES \u002F ZOO (gradient-free estimation)\n\nWhen confidence scores are returned, estimate gradient via finite differences:\n\n```python\nimport numpy as np\n\ndef nes_gradient_estimate(query_fn, x, sigma=0.01, n=50):\n    \"\"\"Natural Evolution Strategy gradient estimate from prediction scores.\"\"\"\n    grads = np.zeros_like(x)\n    for _ in range(n):\n        noise = np.random.randn(*x.shape)\n        pos = query_fn(x + sigma * noise)\n        neg = query_fn(x - sigma * noise)\n        grads += (pos - neg) * noise\n    return grads \u002F (2 * n * sigma)\n\n# query_fn: callable that sends x to the API and returns target-class score\n# Use grads to step: x_adv = x - alpha * np.sign(grads)\n```\n\n### 3b. Transfer attack (surrogate model)\n\nThe reference approach from 13o-bbr-bbq\u002Fmachine_learning_security's CNN_test:\n\n```bash\n# 1. Collect input samples via normal API usage\n# 2. Train a local surrogate on (input, label) pairs\n# 3. Run white-box attack (FGSM\u002FPGD) against surrogate\n# 4. Transfer adversarial examples to the target\n# Transferability is high when surrogate and target share architecture family\n\npip install cleverhans\npython - \u003C\u003C'EOF'\nimport tensorflow as tf\nfrom cleverhans.tf2.attacks.fast_gradient_method import fast_gradient_method\n# Build surrogate_model from collected (x, y) pairs\nx_adv = fast_gradient_method(surrogate_model, x_test, eps=0.03, norm=np.inf)\n# Submit x_adv to the black-box API\nEOF\n```\n\n### 3c. Hard-label: HopSkipJump\n\nWhen only the predicted label is returned (no scores):\n\n```bash\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import HopSkipJump\nfrom art.estimators.classification import BlackBoxClassifier\n\ndef predict_fn(x):\n    # Call your API here, return one-hot array\n    ...\n\nclf = BlackBoxClassifier(predict_fn, input_shape=(32,32,3), nb_classes=10)\nattack = HopSkipJump(classifier=clf, max_iter=50, max_eval=1000)\nx_adv = attack.generate(x=x_test[:5])\nEOF\n```\n\n---\n\n## 4. Physical-world \u002F domain-specific attacks\n\n### 4a. Adversarial patches (object detection, face recognition)\n\nCraft a printable patch that fools YOLO\u002FRetinaNet regardless of placement:\n\n```bash\npip install adversarial-robustness-toolbox\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import AdversarialPatch\nattack = AdversarialPatch(classifier=clf, rotation_max=22.5, scale_min=0.1, scale_max=1.0)\npatch, mask = attack.generate(x=x_train[:100])\n# Print patch, affix to subject — fools real-time camera-based classifiers\nEOF\n```\n\n### 4b. Malware binary perturbation (AV\u002FEDR evasion)\n\nFor ML-based static AV\u002FEDR (MalConv, EMBER model):\n\n```bash\n# Append bytes to DOS overlay — does not affect execution, changes feature vector\npython - \u003C\u003C'EOF'\nimport lief, numpy as np\n\nbinary = lief.parse(\"sample.exe\")\n# Append random payload to avoid changing import table \u002F section headers\nbinary.dos_stub = bytes(np.random.randint(0, 256, 64))\nbuilder = lief.PE.Builder(binary)\nbuilder.build()\nbuilder.write(\"sample_adv.exe\")\n# Iterate: score the modified binary against the target ML model\n# Stop when target-class score drops below threshold\nEOF\n```\n\n### 4c. Network traffic \u002F IDS evasion\n\nFor ML-based IDS (random-forest or neural-net on flow features):\n\n```bash\n# Identify mutable features: packet timing, padding, flow ordering\n# Immutable: payload content that triggers the attack itself\n# Strategy: inflate packet-count, adjust inter-arrival times to move\n# feature vector away from known-malicious region\n\nscapy - \u003C\u003C'EOF'\nfrom scapy.all import *\n# Add benign-looking padding packets to inflate flow duration\n# Shift timing to match benign profile learned from surrogate\nEOF\n```\n\n---\n\n## 5. Tooling reference\n\n| Tool | Attack type | Notes |\n|---|---|---|\n| IBM ART (`adversarial-robustness-toolbox`) | White+black-box, patches, physical | Best all-round; PyTorch\u002FTF\u002FSklearn |\n| CleverHans | White-box FGSM\u002FPGD | TF2 native |\n| Foolbox | White+black-box | NumPy-first, easy API |\n| AutoZOOM | Black-box score-based | Autoencoder-based gradient compression |\n| deep-pwning | Legacy Metasploit-style | Largely unmaintained |\n\n```bash\npip install adversarial-robustness-toolbox foolbox cleverhans\n```\n\n---\n\n## 6. Validation checklist\n\n- [ ] Clean input correctly classified before perturbation (baseline confirmed)\n- [ ] Adversarial input misclassified at the same model endpoint\n- [ ] Perturbation is within declared epsilon-ball (L-inf or L2 norm verified)\n- [ ] For transfer: adversarial success rate on black-box > 30% (non-trivial transfer)\n- [ ] For physical: photograph and re-score — confirm printed patch works under camera\n\n---\n\n## 7. Detection signals (defender perspective)\n\n- Input pixel distribution deviates from natural image statistics (high-frequency noise)\n- Feature-squeeze defence: score changes drastically when input is median-filtered\n- Input preprocessing (JPEG re-encode, resize+crop) breaks gradient-based perturbations\n- Ensemble disagreement: 3 models with different architectures disagree on the adversarial input\n- Activation distribution of adversarial inputs differs from clean examples (mahalanobis-distance detector)\n","# Adversarial ML Evasion\n\nAdversarial examples are inputs modified with small, deliberate perturbations that reliably\ncause a trained classifier to produce incorrect predictions. The perturbation is typically\nimperceptible to humans or functionally irrelevant, while the model's decision boundary is\ncrossed. This is distinct from LLM jailbreaks: the target is a trained discriminative model\n(CNN, gradient-boosted tree, SVM, LSTM) served via an API or embedded in a product.\n\n> **Authorized use only.** All testing must be conducted against systems you own or have\n> explicit written permission to assess. Generating adversarial inputs against production\n> ML APIs without authorization may violate the CFAA and equivalent statutes.\n\n---\n\n## ATT&CK Mapping\n\n| Technique | Use |\n|---|---|\n| T1562.001 — Impair Defenses: Disable or Modify Tools | Evade ML-based AV\u002FEDR\u002FIDS |\n| T1036 — Masquerading | Make malicious content look benign to a classifier |\n| T1027 — Obfuscated Files or Information | Perturb a file\u002Fimage to fool static ML analysis |\n\n---\n\n## 1. Reconnaissance — understand the target model\n\nBefore crafting perturbations, determine the attack surface:\n\n```bash\n# 1a. Identify the ML stack from job postings, open-source repos, error messages\n# Common deployment stacks: TensorFlow Serving, TorchServe, ONNX Runtime, SageMaker, AzureML\n\n# 1b. Probe input shape and output format\ncurl -s -X POST \"$TARGET\u002Fpredict\" \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\"data\": [[0.0]*784]}' | jq .\n# Note: number of output classes, confidence scores vs hard labels\n# Confidence scores = white-box-equivalent gradient signal via finite differences\n\n# 1c. Check if the API returns confidence values (score-based black-box) or label only (hard-label)\n# Score-based: enables gradient estimation, ZOO, NES attacks\n# Label-only: requires hard-label attacks (HopSkipJump, QEBA)\n```\n\n---\n\n## 2. White-box evasion (full model access)\n\nUse when you have the model weights (pentest scope, internal system, open-source model).\n\n### 2a. FGSM — Fast Gradient Sign Method (single step)\n\n```python\nimport torch, torchvision.transforms as T\nfrom PIL import Image\nimport requests, json\n\ndef fgsm(model, x, y_true, epsilon=0.03):\n    \"\"\"Single-step gradient attack. Fast, low distortion for strong models.\"\"\"\n    x.requires_grad_(True)\n    loss = torch.nn.CrossEntropyLoss()(model(x), y_true)\n    loss.backward()\n    return (x + epsilon * x.grad.sign()).clamp(0, 1).detach()\n```\n\n### 2b. PGD — Projected Gradient Descent (iterative, stronger)\n\n```python\ndef pgd(model, x, y_true, epsilon=0.03, alpha=0.007, steps=40):\n    \"\"\"Madry's PGD — iterative FGSM with projection back into epsilon-ball.\"\"\"\n    x_adv = x.clone().detach().requires_grad_(True)\n    for _ in range(steps):\n        loss = torch.nn.CrossEntropyLoss()(model(x_adv), y_true)\n        loss.backward()\n        with torch.no_grad():\n            x_adv = x_adv + alpha * x_adv.grad.sign()\n            delta = torch.clamp(x_adv - x, -epsilon, epsilon)\n            x_adv = torch.clamp(x + delta, 0, 1).detach().requires_grad_(True)\n    return x_adv\n```\n\n### 2c. C&W — Carlini & Wagner (optimization-based, minimal distortion)\n\n```bash\npip install adversarial-robustness-toolbox  # IBM ART\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import CarliniL2Method\nfrom art.estimators.classification import PyTorchClassifier\n# wrap your model in PyTorchClassifier, then:\nattack = CarliniL2Method(classifier=clf, confidence=0.5, max_iter=100)\nx_adv = attack.generate(x=x_test[:10])\nEOF\n```\n\n---\n\n## 3. Black-box evasion (API access only)\n\n### 3a. Score-based: NES \u002F ZOO (gradient-free estimation)\n\nWhen confidence scores are returned, estimate gradient via finite differences:\n\n```python\nimport numpy as np\n\ndef nes_gradient_estimate(query_fn, x, sigma=0.01, n=50):\n    \"\"\"Natural Evolution Strategy gradient estimate from prediction scores.\"\"\"\n    grads = np.zeros_like(x)\n    for _ in range(n):\n        noise = np.random.randn(*x.shape)\n        pos = query_fn(x + sigma * noise)\n        neg = query_fn(x - sigma * noise)\n        grads += (pos - neg) * noise\n    return grads \u002F (2 * n * sigma)\n\n# query_fn: callable that sends x to the API and returns target-class score\n# Use grads to step: x_adv = x - alpha * np.sign(grads)\n```\n\n### 3b. Transfer attack (surrogate model)\n\nThe reference approach from 13o-bbr-bbq\u002Fmachine_learning_security's CNN_test:\n\n```bash\n# 1. Collect input samples via normal API usage\n# 2. Train a local surrogate on (input, label) pairs\n# 3. Run white-box attack (FGSM\u002FPGD) against surrogate\n# 4. Transfer adversarial examples to the target\n# Transferability is high when surrogate and target share architecture family\n\npip install cleverhans\npython - \u003C\u003C'EOF'\nimport tensorflow as tf\nfrom cleverhans.tf2.attacks.fast_gradient_method import fast_gradient_method\n# Build surrogate_model from collected (x, y) pairs\nx_adv = fast_gradient_method(surrogate_model, x_test, eps=0.03, norm=np.inf)\n# Submit x_adv to the black-box API\nEOF\n```\n\n### 3c. Hard-label: HopSkipJump\n\nWhen only the predicted label is returned (no scores):\n\n```bash\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import HopSkipJump\nfrom art.estimators.classification import BlackBoxClassifier\n\ndef predict_fn(x):\n    # Call your API here, return one-hot array\n    ...\n\nclf = BlackBoxClassifier(predict_fn, input_shape=(32,32,3), nb_classes=10)\nattack = HopSkipJump(classifier=clf, max_iter=50, max_eval=1000)\nx_adv = attack.generate(x=x_test[:5])\nEOF\n```\n\n---\n\n## 4. Physical-world \u002F domain-specific attacks\n\n### 4a. Adversarial patches (object detection, face recognition)\n\nCraft a printable patch that fools YOLO\u002FRetinaNet regardless of placement:\n\n```bash\npip install adversarial-robustness-toolbox\npython - \u003C\u003C'EOF'\nfrom art.attacks.evasion import AdversarialPatch\nattack = AdversarialPatch(classifier=clf, rotation_max=22.5, scale_min=0.1, scale_max=1.0)\npatch, mask = attack.generate(x=x_train[:100])\n# Print patch, affix to subject — fools real-time camera-based classifiers\nEOF\n```\n\n### 4b. Malware binary perturbation (AV\u002FEDR evasion)\n\nFor ML-based static AV\u002FEDR (MalConv, EMBER model):\n\n```bash\n# Append bytes to DOS overlay — does not affect execution, changes feature vector\npython - \u003C\u003C'EOF'\nimport lief, numpy as np\n\nbinary = lief.parse(\"sample.exe\")\n# Append random payload to avoid changing import table \u002F section headers\nbinary.dos_stub = bytes(np.random.randint(0, 256, 64))\nbuilder = lief.PE.Builder(binary)\nbuilder.build()\nbuilder.write(\"sample_adv.exe\")\n# Iterate: score the modified binary against the target ML model\n# Stop when target-class score drops below threshold\nEOF\n```\n\n### 4c. Network traffic \u002F IDS evasion\n\nFor ML-based IDS (random-forest or neural-net on flow features):\n\n```bash\n# Identify mutable features: packet timing, padding, flow ordering\n# Immutable: payload content that triggers the attack itself\n# Strategy: inflate packet-count, adjust inter-arrival times to move\n# feature vector away from known-malicious region\n\nscapy - \u003C\u003C'EOF'\nfrom scapy.all import *\n# Add benign-looking padding packets to inflate flow duration\n# Shift timing to match benign profile learned from surrogate\nEOF\n```\n\n---\n\n## 5. Tooling reference\n\n| Tool | Attack type | Notes |\n|---|---|---|\n| IBM ART (`adversarial-robustness-toolbox`) | White+black-box, patches, physical | Best all-round; PyTorch\u002FTF\u002FSklearn |\n| CleverHans | White-box FGSM\u002FPGD | TF2 native |\n| Foolbox | White+black-box | NumPy-first, easy API |\n| AutoZOOM | Black-box score-based | Autoencoder-based gradient compression |\n| deep-pwning | Legacy Metasploit-style | Largely unmaintained |\n\n```bash\npip install adversarial-robustness-toolbox foolbox cleverhans\n```\n\n---\n\n## 6. Validation checklist\n\n- [ ] Clean input correctly classified before perturbation (baseline confirmed)\n- [ ] Adversarial input misclassified at the same model endpoint\n- [ ] Perturbation is within declared epsilon-ball (L-inf or L2 norm verified)\n- [ ] For transfer: adversarial success rate on black-box > 30% (non-trivial transfer)\n- [ ] For physical: photograph and re-score — confirm printed patch works under camera\n\n---\n\n## 7. Detection signals (defender perspective)\n\n- Input pixel distribution deviates from natural image statistics (high-frequency noise)\n- Feature-squeeze defence: score changes drastically when input is median-filtered\n- Input preprocessing (JPEG re-encode, resize+crop) breaks gradient-based perturbations\n- Ensemble disagreement: 3 models with different architectures disagree on the adversarial input\n- Activation distribution of adversarial inputs differs from clean examples (mahalanobis-distance detector)\n",{"name":5,"metadata":23,"description":7,"allowed-tools":28},{"tags":24,"subdomain":25,"when_to_use":26,"mitre_attack":27},"adversarial-ml, evasion, classifier-bypass, computer-vision, malware-evasion, ids-evasion","ai-security","adversarial example evasion classifier bypass ml model fool misclassify image recognition malware detector IDS antivirus neural network perturbation FGSM PGD","T1562.001, T1036, T1027","Bash Read Write","main","376de63e2b186eeb9856476ae89186e62b1d3fbe","packages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fanalyst\u002Fadversarial-ml-evasion","https:\u002F\u002Fgithub.com\u002Fdnzengou\u002Fevo-decepticon","https:\u002F\u002Fgithub.com\u002Fdnzengou","2026-11-09 16:34:40",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add dnzengou\u002Fevo-decepticon --skill adversarial-ml-evasion",{"tool":41,"command":42},"bunx","bunx skills add dnzengou\u002Fevo-decepticon --skill adversarial-ml-evasion",{"tool":44,"command":45},"pnpm","pnpm dlx skills add dnzengou\u002Fevo-decepticon --skill adversarial-ml-evasion","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fanalyst\u002Fadversarial-ml-evasion\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":18},"01M78JAX15GZR5Q5S7RMPWFX43","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":18},"01M78JARXE0ZER8P81SFD024YP","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":18},"01M78JANJNH8VMR0CK1Y6AM74G","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":18},"01M78JAFCYR2BAJF7KH0ZCXREA","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":18},"01M78JAAQ5T08FR9MY5SFNRWBM","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\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":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":18},"01M78JA6799JQZE0M9B21HRMGY","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\u002Fdnzengou\u002Fevo-decepticon\u002Fblob\u002Fmain\u002Fpackages\u002Fdecepticon\u002Fdecepticon\u002Fskills\u002Fstandard\u002Fwireless\u002Fevil-twin-karma\u002FSKILL.md",[],[]]