-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathAI_Attack_Atlas_v2_expanded.opml
More file actions
2656 lines (959 loc) · 75.9 KB
/
AI_Attack_Atlas_v2_expanded.opml
File metadata and controls
2656 lines (959 loc) · 75.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<?xml version="1.0" encoding="UTF-8"?>
<opml version="2.0">
<head>
<title>AI Attack Atlas (Operator View)</title>
<dateCreated>2025-08-25T21:56:49.153484Z</dateCreated>
</head>
<body>
<outline text="AI Attack Atlas (Operator View): Preconditions → Enumeration → Exploit → Pivot → Goals">
<outline text="Goal States (checkpoints & pivots)">
<outline text="Model-weight exfiltration (foundation / fine-tuned)"/>
<outline text="Sensitive training data recovery (exact strings, PII, secrets)"/>
<outline text="Cross-tenant data access / data sovereignty violation"/>
<outline text="Arbitrary tool/action execution via agent control"/>
<outline text="Persistent poisoning (dataset / KB / embeddings / reward model / template)"/>
<outline text="Undetected model/inference tamper (silent bias / backdoor / jailbreak persistence)"/>
<outline text="Cloud account escalation (provider / hosting / MLOps / K8s)"/>
<outline text="Supply-chain compromise (models, datasets, packages, plugins, MCP tools)"/>
<outline text="GPU/TEE side-channel exfiltration"/>
<outline text="Monetary burn (DoS/Auto-scaling token drain, quota exhaustion)"/>
<outline text="Federated learning participant compromise / global model corruption"/>
<outline text="Edge device model theft / reverse engineering"/>
<outline text="Reinforcement learning reward manipulation / policy corruption"/>
<outline text="AutoML search space poisoning / architecture backdoors"/>
<outline text="Synthetic data generator compromise / distribution shift attacks"/>
</outline>
<outline text="Initial Preconditions (select what you have)">
<outline text="Black-box chat access only (no plugins/tools)">
<outline text="Direct prompt injection / jailbreaks (policy evasion, content policy bypass)"/>
<outline text="Indirection via content ingestion (web pages, PDFs, emails) → Indirect prompt injection"/>
<outline text="Hallucination steering for data exfil (model invents but also reveals indexed content)"/>
<outline text="Model extraction surface (query synthesis for surrogate training; confidence-less extraction)"/>
<outline text="Model fingerprinting (provider, family, version inference via quirks)"/>
</outline>
<outline text="Chat with tools / function calling (JSON schema based)">
<outline text="Tool selection hijack (description/param priming)"/>
<outline text="Argument smuggling in natural language fields (free-form → JSON args)"/>
<outline text="Return-channel injection (tool output → appended to conversation)"/>
<outline text="Multi-step tool loops (planner/executor) feedback poisoning"/>
</outline>
<outline text="RAG application (vector DB + KB)">
<outline text="KB write access (wiki, CMS, SharePoint, Git, S3) → payload planting"/>
<outline text="Embedding pipeline manipulation (chunking, metadata, collection mix-up)"/>
<outline text="Index poisoning (topic hijack, anchor hijack, negative retrieval)"/>
<outline text="Citation spoofing & retrieval coercion (rare-token anchors, adversarial strings)"/>
</outline>
<outline text="Agentic system (autonomous/looped)">
<outline text="Goal re-specification via hidden instructions"/>
<outline text="Tool graph/skill registry poisoning (MCP/tool schemas, plugin manifests)"/>
<outline text="Memory poisoning (episodic/scratchpad) → long-term drift"/>
<outline text="Output-to-input loop abuse (self-referential notes become commands)"/>
</outline>
<outline text="Fine-tune or SFT rights (no base weights)">
<outline text="Prompt/response set poisoning (backdoors, trigger phrases)"/>
<outline text="Safety inversion via few-shot anchors"/>
<outline text="PII lubricant examples increasing leakage propensity"/>
</outline>
<outline text="RLHF/Preference data access">
<outline text="Preference dataset poisoning (sentiment flips, reward hacking)"/>
<outline text="Reward model trigger backdoors (keyword / token triggers)"/>
<outline text="Long-output cost-inflation poisoning (token-bloat preference)"/>
</outline>
<outline text="Foundation model weights access (read)">
<outline text="Parameter-space watermark removal / backdoor injection (diffusion/LLM)"/>
<outline text="Layer surgery (LoRA/adapter grafts)"/>
<outline text="Quantization/repacking time-of-use tamper"/>
</outline>
<outline text="Inference server access (host/container)">
<outline text="Token/secret scraping (env, config, logs)"/>
<outline text="Model file interception (volume mounts, NFS, object storage)"/>
<outline text="Side-channel enabling (GPU/local mem, compression channels)"/>
<outline text="Response cache scraping (kv caches, eval caches)"/>
</outline>
<outline text="MLOps platform access (MLflow/Kubeflow/Vertex/SageMaker)">
<outline text="Model registry poisoning (malicious pickle/artifacts)"/>
<outline text="Pipeline spec tamper (YAML/DSL → supply-chain)"/>
<outline text="UI/XSS to session hijack → secret theft"/>
<outline text="CI/CD runner compromise → cluster pivot"/>
</outline>
<outline text="Cloud role/service principal/API key">
<outline text="Secrets manager enumeration → downstream creds"/>
<outline text="Storage buckets of training/embeddings/weights"/>
<outline text="GPU fleet metadata access → co-tenant positioning"/>
<outline text="Function-as-a-Service hooks to exfil / run"/>
</outline>
<outline text="Neighbor GPU/co-tenant (shared node)">
<outline text="Local memory residue reading (LeftoverLocals-class)"/>
<outline text="GPU compression side-channels (GPU.zip-class)"/>
<outline text="PCIe/NUMA topology inference → cross-container leakage risk"/>
</outline>
<outline text="Training data partial access">
<outline text="Targeted canary detection / canary extraction"/>
<outline text="Backdoor seeding (trigger-token behaviors)"/>
<outline text="Membership inference ground-truth construction"/>
</outline>
<outline text="Vector DB read-only">
<outline text="Embedding inversion risk analysis (PII recovery feasibility)"/>
<outline text="Nearest-neighbor probing to infer sensitive clusters"/>
</outline>
<outline text="Plugin/extension publishing rights (marketplaces/IDE)">
<outline text="Agent capability backdooring (over-privileged actions)"/>
<outline text="Update-channel hijack (silent manifest upgrades)"/>
</outline>
<outline text="Reinforcement Learning environment access">
<outline text="Environment specification tampering (state/action space manipulation)"/>
<outline text="Reward signal interception and modification"/>
<outline text="Adversarial state injection during training"/>
<outline text="Policy network direct access for trojan insertion"/>
</outline>
<outline text="Federated Learning participant role">
<outline text="Local model update poisoning (Byzantine attacks)"/>
<outline text="Gradient manipulation and amplification"/>
<outline text="Client selection gaming / Sybil participation"/>
<outline text="Privacy attack coordination across participants"/>
</outline>
<outline text="Edge device / IoT deployment access">
<outline text="Model extraction via timing/power analysis"/>
<outline text="Hardware fault injection (rowhammer, glitching)"/>
<outline text="Firmware backdooring for persistent access"/>
<outline text="OTA update channel compromise"/>
</outline>
<outline text="AutoML / NAS platform access">
<outline text="Search space constraint manipulation"/>
<outline text="Architecture evaluation metric poisoning"/>
<outline text="Supernetwork weight corruption"/>
<outline text="Hardware-aware NAS targeting specific accelerators"/>
</outline>
<outline text="Multi-modal system interface">
<outline text="Cross-modal injection (image→text, audio→vision)"/>
<outline text="Modality alignment poisoning"/>
<outline text="Fusion layer manipulation"/>
<outline text="Attention mechanism hijacking across modalities"/>
</outline>
</outline>
<outline text="Cross-cutting Attack Classes (what can be done)">
<outline text="Prompt/Policy Attacks">
<outline text="Direct jailbreaks (role confusion, multi-turn dilution)"/>
<outline text="Indirect prompt injection (in-document/web/payload)"/>
<outline text="Multi-agent propagation (self-referential, worm-like patterns)"/>
<outline text="Insecure Output Handling (downstream code/HTML/SQL/URL execution)"/>
<outline text="Tool-call schema hijack (name/desc/param poisoning)"/>
</outline>
<outline text="Data Poisoning">
<outline text="Pretraining/web-scale seed poisoning (corpus hotspots, robots.txt-ignoring)"/>
<outline text="SFT/fine-tune poisoning (style transfer triggers, harmful response bias)"/>
<outline text="Reward-model poisoning (preference flips, token-length bloat)"/>
<outline text="RAG KB poisoning (anchoring, doc stuffing, prompt-in-doc)"/>
<outline text="Diffusion concept poisoning (Nightshade-style prompt-specific)"/>
<outline text="Federated learning update poisoning (model replacement, gradient scaling)"/>
<outline text="Reinforcement learning trajectory poisoning"/>
<outline text="Synthetic data distribution poisoning"/>
<outline text="Active learning query manipulation"/>
<outline text="Continual learning memory poisoning"/>
</outline>
<outline text="Privacy/Model Theft">
<outline text="Model extraction via black-box queries (surrogate fidelity)"/>
<outline text="Training data extraction (verbatim memorization mining)"/>
<outline text="Membership inference (is X in training?)"/>
<outline text="Model inversion (recover inputs/attributes from outputs)"/>
<outline text="Embedding inversion (vec2text; reconstructing source text)"/>
<outline text="Federated learning participant data reconstruction"/>
<outline text="Gradient inversion attacks"/>
<outline text="Architecture extraction via API timing"/>
<outline text="Hyperparameter inference attacks"/>
<outline text="Differential privacy budget attacks"/>
</outline>
<outline text="Adversarial ML (inputs)">
<outline text="Additive perturbations (FGSM/PGD)"/>
<outline text="Physical-world patches/objects (stickers, glasses)"/>
<outline text="Universal perturbations (image-agnostic)"/>
<outline text="Audio/ultrasonic/laser command injection"/>
<outline text="Time-series & tabular adversarials (finance, ICS)"/>
<outline text="Semantic adversarial examples (maintaining meaning)"/>
<outline text="Transferable adversarial examples"/>
<outline text="Query-efficient black-box attacks"/>
<outline text="Adversarial reprogramming"/>
<outline text="Backdoor triggers in inputs"/>
</outline>
<outline text="Supply Chain & Platform">
<outline text="Malicious models/artifacts (pickle, model cards, weights)"/>
<outline text="Package ecosystem abuse (pip/npm/conda)"/>
<outline text="MLOps UI/API vulns (XSS, CSRF, deserialization)"/>
<outline text="K8s/KFP/CI runners (privilege escalation)"/>
<outline text="Model hub dependency confusion / typosquatting"/>
<outline text="Dataset poisoning at source"/>
<outline text="Pretrained model backdoors"/>
<outline text="Framework vulnerability exploitation"/>
<outline text="Hardware supply chain attacks"/>
<outline text="Cloud provider API abuse"/>
</outline>
<outline text="Side-channels & Isolation">
<outline text="GPU local memory residue leakage"/>
<outline text="GPU compression side-channel"/>
<outline text="Speculative execution & cache timing"/>
<outline text="Prompt/token caches & logs"/>
<outline text="TEE/trust boundary misconfigurations"/>
<outline text="Power analysis attacks"/>
<outline text="Electromagnetic emanation analysis"/>
<outline text="Acoustic cryptanalysis"/>
<outline text="Temperature-based side channels"/>
<outline text="Network timing analysis"/>
</outline>
<outline text="Economic & Availability">
<outline text="Token/compute DoS (junk prompts, recursive tasks)"/>
<outline text="Context-bloat amplification (RAG over-retrieval)"/>
<outline text="Eval & guardrail abuse to raise cost"/>
<outline text="Quota/key draining via hidden loops"/>
<outline text="Sponge attacks (maximizing energy consumption)"/>
<outline text="API rate limit exploitation"/>
<outline text="Batch processing attacks"/>
<outline text="Resource starvation via model complexity"/>
<outline text="Cascading failure induction"/>
</outline>
<outline text="Explainability & Interpretability Attacks">
<outline text="Explanation manipulation (LIME/SHAP gaming)"/>
<outline text="Saliency map deception"/>
<outline text="Feature importance masking"/>
<outline text="Counterfactual explanation attacks"/>
<outline text="Attribution method exploitation"/>
</outline>
<outline text="Meta-Learning & Few-Shot Attacks">
<outline text="Support set poisoning"/>
<outline text="Task distribution manipulation"/>
<outline text="Meta-parameter corruption"/>
<outline text="Few-shot backdoor injection"/>
<outline text="Cross-task interference attacks"/>
</outline>
</outline>
<outline text="Paradigm-Specific Paths (deep)">
<outline text="LLMs (chat/completion)">
<outline text="Enumeration">
<outline text="Model/provider/version & safety profile"/>
<outline text="Tooling enabled? (functions/tools/plugins)"/>
<outline text="Context sources (web/RAG/KB/memory)"/>
<outline text="Output sinks (renderers, shells, DBs, endpoints)"/>
</outline>
<outline text="Attack vectors">
<outline text="Direct/indirect prompt injection; multi-turn framing"/>
<outline text="System prompt scraping via shadow-token hints"/>
<outline text="Context poisoning via citations/footnotes"/>
<outline text="JSON argument smuggling; type confusion"/>
<outline text="Return-channel injection to reframe next turn"/>
</outline>
<outline text="Pivots">
<outline text="Tool execution capability gains"/>
<outline text="Data exfiltration via browsing/downloader"/>
<outline text="Policy collapse → unrestricted code suggestion"/>
<outline text="Key/secret exposure via logs/error channels"/>
</outline>
</outline>
<outline text="RAG Systems (retrieval-augmented)">
<outline text="Enumeration">
<outline text="Index type (HNSW/IVF), chunking, metadata"/>
<outline text="Write paths (CMS/wiki/fileshare/git/S3)"/>
<outline text="Pre-processing (OCR, splitters, parsers)"/>
<outline text="Citations & grounding enforcement"/>
</outline>
<outline text="Attack vectors">
<outline text="KB poisoning (payload-in-doc; hidden text)"/>
<outline text="Anchor hijack (rare-token, unicode confusables)"/>
<outline text="Metadata skew (priority, freshness, source)"/>
<outline text="Embedding collision crafting; adversarial queries"/>
<outline text="Citation spoofing/URL override"/>
</outline>
<outline text="Pivots">
<outline text="Persistent agent instruction via KB"/>
<outline text="Cross-collection bleed into other apps"/>
<outline text="Embedding theft → inversion risk (PII)"/>
</outline>
</outline>
<outline text="Agentic AI (planners/executors/memory)">
<outline text="Enumeration">
<outline text="Planner vs tool executor separation"/>
<outline text="Tool registry (MCP/JSON schema) provenance"/>
<outline text="Memory stores (short/long-term)"/>
<outline text="Safety sandboxes (browser/fs/exec)"/>
</outline>
<outline text="Attack vectors">
<outline text="Schema poisoning (name/desc/param/defaults)"/>
<outline text="Advanced tool poisoning (crafted tool outputs)"/>
<outline text="Plan injection (goal rewrite, recursive self-calls)"/>
<outline text="Multi-agent propagation (worm-like payloads)"/>
<outline text="LLM→Tool Output Injection (model-crafted prompts to tools)">
<outline text="Downstream sink mapping (shell/SQL/NoSQL/JS/URL/Filesystem)"/>
<outline text="Context-splitting attacks (tool expects JSON; model emits free text wrapper)"/>
<outline text="ID binding confusion (mismatched tool_call_id → wrong response consumed)"/>
<outline text="Tool-return reflection (error strings become next-step instructions)"/>
</outline>
<outline text="JSON/Schema Type Confusion">
<outline text="additionalProperties wildcard abuse"/>
<outline text="Enum fallthrough (unknown → default dangerous action)"/>
<outline text="Required-but-null coercion (validator bugs)"/>
<outline text="Injected URLs/paths leading to SSRF/LFI patterns"/>
</outline>
<outline text="Planner/Executor Divergence">
<outline text="Plan stuffing (hidden subtasks with high priority)"/>
<outline text="Self-referencing memories (notes → commands)"/>
<outline text="Skill name collisions (benign vs malicious homographs)"/>
</outline>
</outline>
<outline text="Pivots">
<outline text="Arbitrary action execution chain"/>
<outline text="Credential use through tools"/>
<outline text="Outbound spam/data exfil loops"/>
</outline>
</outline>
<outline text="Fine-tuning / SFT">
<outline text="Enumeration">
<outline text="Dataset sources & provenance"/>
<outline text="PII presence, deduplication, canaries"/>
<outline text="Safety filters on outputs during SFT"/>
</outline>
<outline text="Attack vectors">
<outline text="Trigger-based backdoors in examples"/>
<outline text="Stylistic jailbreak primers"/>
<outline text="Response leakage amplifiers (template bias)"/>
</outline>
<outline text="Pivots">
<outline text="Backdoor activation in production"/>
<outline text="Model card/watermark inconsistencies"/>
</outline>
</outline>
<outline text="RLHF / Reward Models">
<outline text="Enumeration">
<outline text="Preference pipeline (pairwise, listwise)"/>
<outline text="Public datasets mixed-in?"/>
<outline text="Reward model architecture & triggers"/>
</outline>
<outline text="Attack vectors">
<outline text="Preference poisoning (sentiment/length bias)"/>
<outline text="Reward hacking/wireheading tendencies"/>
<outline text="Trigger-token reward spikes"/>
</outline>
<outline text="Pivots">
<outline text="Global behavior drift (longer outputs, policy changes)"/>
<outline text="Residual vulnerabilities despite guardrails"/>
</outline>
</outline>
<outline text="Vision (classification/detection/segmentation)">
<outline text="Attack vectors">
<outline text="FGSM/PGD; universal perturbations"/>
<outline text="Physical-world adversarial patches"/>
<outline text="Backdoor triggers in datasets"/>
<outline text="Data labeling attacks"/>
</outline>
<outline text="Pivots">
<outline text="Safety-critical misclassification (AV/industrial)"/>