CVE-2025-1194:transformers: Regular-expression denial of service
A reconstruction of the flaw, and the evidence path Lachesis follows to make it visible.
A long run of digits and commas (no trailing currency unit) fed through the tokenizer's normalize step.
That is the shape of the failure. The rest of this case file follows the evidence behind it.
What was known.
What Lachesis established.
Known before Lachesis ran
The CVE and vulnerable release were selected from public history. We are not claiming novel discovery.
- CVE
- CVE-2025-1194
- Vulnerable target
- v4.49.0 (affected: <4.50.0) · vulnerable
- Prior knowledge
- CVE-2025-1194 · v4.49.0 (affected: <4.50.0) is vulnerable
Independent re-detection
The experiment graph was built around the vulnerable release, enriched (pass2 dataflow + atropos bind + pass3 semantic), then enumerated BLIND over the whole constructor registry (no family hard-coded, no target line seeded). The lead below is the class-matched candidate the enumerator surfaced on its own; the provenance and body reads are the tool's own follow-up operations. Prose in `adjudication`/`trace` is MY reading of the verbatim source + official fix, clearly separated from tool output.
- Seeded inputs
- None
- Run timestamp
- 2026-08-29T00:00:00Z
- Evidence artifact
- cve-2025-1194.kuzu (built from the vulnerable release, enriched pass2+pass3; graph + sidecars deleted after capture)
The commands and outputs below come from this recorded Lachesis run.
What Lachesis reconstructed.
One historical repository. Four captured queries. A complete source-to-sink argument.
Lachesis follows the same evidence path through unfamiliar repositories.
Here is how the path becomes visible.
These are the recorded questions Lachesis asked of the historical vulnerable code, followed by the raw result and source location each query returned.
Captured runA real replay of Lachesis over the graph we built for this case. Every `output` below is verbatim captured tool text; the graph was built bounded (hard memory/time cap) and released immediately after capture, so nothing accumulates.
Load the graph and list every bug family
Load the graph built from the vulnerable release and list every bug family the blind enumerator knows -- no family is pre-selected.
{"move": "load_graph", "graph": "~/.lachesis/graphs/cve-2025-1194.kuzu", "profile": "all", "nodes": 9774}
CANDIDATE_CENSUS
move: candidate_census
taxonomy (8):
domain=lifecycle title=Resource lifecycle meaning=an operation that acquires, releases, uses, or transfers a tracked resource obligation=resource operations stay within the object's lifetime primary=True languages=['c', 'python', 'javascript', 'typescript'] enumerable=True families=[{'family': 'acquire', 'obligation': 'resource acquisition', 'constructor': 'lifecycle.acquire', 'enumerable': True}, {'family': 'release', 'obligation': 'resource release', 'constructor': 'lifecycle.release', 'enumerable': True}, {'family': 'use', 'obligation': 'tracked resource use', 'constructor': 'lifecycle.use', 'enumerable': True}, {'family': 'escape', 'obligation': 'resource ownership transfer', 'constructor': 'lifecycle.escape', 'enumerable': True}, {'family': 'leak', 'obligation': 'Allocated object remains live at an exit', 'constructor'
... [truncated]A small line with a large consequence.
no linear-time/possessive formulation; the compiled regex admits exponential backtracking on adversarial digit runs
self.content_repatter6 = re.compile(r"((0|[1-9]\d*|[1-9]\d{0,2}(,\d{3})+)*億)*...")// 4.50.0 (commit 92c5ca9) rewrote content_repatter6 with possessive quantifiers / atomic groups (3.11+) and flattened alternations (<3.11) so no quantified group overlaps its neighbour
The original record.
Read the historical advisory and vulnerability record behind this reconstruction.