Correct eleven MITRE ATLAS technique citations

[astrogilda]

Summary

Eleven MITRE ATLAS control mappings pair an identifier with a name that ATLAS does not give it, across 48 rows in the three ATLAS tables. In six the identifier is right and the label is superseded; in five the label is right and current but belongs to a different technique. Every target is checked against the pinned ATLAS release named below. Judgement calls are kept out of this pull request and raised in #60.

DSGAI mapping

All twenty-one DSGAI entries, plus ASI03, ASI07, ASI08 and LLM10, which is 25 entry files, reached by regenerating from the three corrected markdown tables.

Checklist

• Data is anonymized (no PII, credentials, or org names)
• Entries reference at least one DSGAI ID where applicable
• Validation scripts pass (python run_all_checks.py) — I could not tick this honestly. data_validation/run_all_checks.py currently prints "Validators not yet implemented - see validators/ for contribution guide" and exits without running a check, so a pass from it would not mean anything. npm run generate succeeds on this branch. npm run build is generate followed by validate and it exits 1, with 51 errors, 67 warnings and 240 passed. That is the same figure npm run validate gives on main at 2aa707a, so the failure is the pre-existing one noted at the end of this description and nothing in this change moves it. Please tell me if there is another gate you would like run.
• One entry per file for dataset contributions — not applicable; this touches crosswalk/ only and adds no dataset records.

---

Eleven MITRE ATLAS control mappings pair an identifier with a name that ATLAS does not give it. In six the identifier is right and the label is superseded, mostly from the pass that moved the matrix off "ML" wording. In five the label is right and current but belongs to a different technique, so the identifier is what moves. This changes 48 rows across the three ATLAS tables and nothing else.

I have kept this to corrections where the primary settles the question on its own, so there should be nothing here to weigh up. Anything needing a judgement call is in the companion issue instead, including two mappings I think are wrong but where the fix could reasonably go either way.

Every target was checked against https://atlas.mitre.org/atlas-data/dist/v6/ATLAS-2026.07.yaml, sha256 0e07bb07fc6423d72cdf24ddc2038a6905bcbc00ba571064153119ee1a5888d4. The dist/ATLAS.yaml one level up in the same distribution directory declares itself deprecated on its first line, and the STIX JSON published at mitre-atlas/atlas-navigator-data lags the YAML: on 3 August it still gave AML.T0020 as "Poison Training Data" and omitted eleven identifiers, so the pinned v6 file is the one to read.

Renames, where the identifier is right and the label moved:

Identifier	Recorded here	ATLAS 2026.07
AML.T0020	Poison Training Data	Training Data Poisoning
AML.T0029	Denial of ML Service	Denial of AI Service
AML.T0013	Discover ML Model Ontology	Discover AI Model Ontology
AML.T0018	Backdoor ML Model	Manipulate AI Model
AML.T0025	Exfiltrate via Cyber Means	Exfiltration via Cyber Means
AML.T0024.000	Membership Inference	Infer Training Data Membership

Identifier errors, where the label is right and points at the wrong technique:

Recorded	Should be	Because
AML.T0057 "Data from Information Repositories"	AML.T0036	AML.T0057 is "LLM Data Leakage"
AML.T0051 "Exploit Public-Facing Application"	AML.T0049	AML.T0051 is "LLM Prompt Injection"
AML.T0031 "Craft Adversarial Data"	AML.T0043	AML.T0031 is "Erode AI Model Integrity"
AML.T0016 "Exfiltration via AI Inference API"	AML.T0024	AML.T0016 is Obtain Capabilities
AML.T0022 "Valid Accounts"	AML.T0012	AML.T0022 appears in neither release I checked

Mostly this is not version drift, and it is worth separating the one case where it is. I read the current v6 release and the deprecated 5.6.0 line. Ten of the eleven pairings match neither, so they were never right. The exception is AML.T0020, which really was "Poison Training Data" in the 5.6.0 line and became "Training Data Poisoning", so that row was correct when it was written. I have not read the intermediate v6 releases, so I can say these do not match the current or the deprecated file, and not that they never appeared anywhere.

AML.T0022 is the odd one out in the other direction: it is absent from both files, so the row cites an identifier no release I read defines. "Valid Accounts" is AML.T0012, which is what the rest of the row describes.

The corrections are applied to the markdown tables rather than to data/entries/, since scripts/generate.js regenerates the entries from the tables and would otherwise revert them. Running npm run generate afterwards propagates the change to 25 entry files: 23 control_id values, 25 control_name values and 4 url values, 52 lines in all. I have included the regenerated entries and bundles so the tree is consistent, and I am happy to drop the docs/*.js bundles from the diff if you would rather regenerate those yourself.

One thing worth flagging that is not a correction. I found these by writing a script that reads each framework's primary source and compares every citation to it, and I would like to offer that separately if it is useful — it is in the companion issue rather than here, because it is a new file and a workflow, and that is a different conversation from fixing eleven rows. Two notes from building it that stand on their own regardless: scripts/generate.js stamps the current date into the four generated bundles, which means a reproducibility check on them can never pass, and npm run validate currently reports 51 errors on main independently of anything in this change.

Edited 2026-08-20 to correct three things in this description. DSGAI21 is changed like every other DSGAI entry, so the exception clause was wrong while the count of 25 was right. The entry regeneration is 52 lines and not 76. And npm run build does not succeed here, because validate fails on the 51 pre-existing errors it also fails on against main.