Segfault and data corruption in tensorflow-lite
Published Sep 25, 2020 · Updated Nov 3, 2026
Original summary
Normalized public record · Updated Nov 3, 2026
Segfault and data corruption in tensorflow-lite
Publisher and source text is retained separately from QFinch interpretation.
QFinch assessment
To mimic Python's indexing with negative values, TFLite uses `ResolveAxis` to convert negative values to positive indices. However, the only check that the converted index is now valid is only present in debug builds: https://github.com/tensorflow/tensorflow/blob/0e68f4d3295eb0281a517c3662f6698992b7b2cf/tensorflow/lite/kernels/internal/reference/reduce.h#L68-L72
Potential impact
To mimic Python's indexing with negative values, TFLite uses ResolveAxis to convert negative values to positive indices. However, the only check that the converted index is now valid is only present in debug builds: https://github.com/tensorflow/tensorflow/blob/0e68f4d3295eb0281a517c3662f6698992b7b2cf/tensorflow/lite/kernels/internal/reference/reduce.hL68-L72 If the DCHECK does not trigger, then code execution moves ahead with a negative index. This, in turn, results in accessing data out of bounds which results in segfaults and/or data corruption.
Evidence: public source recordsAffected products
| Affected package, product or range | Fix status |
|---|---|
| Consult the referenced vendor advisory for affected versions. | Consult source advisory |
Matching a version is only one part of assessing exposure. QFinch has not assessed your environment on this public page.
Remediation
Apply the vendor-supported fixed release or mitigation, then verify the affected component is no longer reachable.
- Confirm the affected product and reachable code path exist in your environment.
- Apply the vendor-supported patch or mitigation from the primary advisory.
- Rescan the exact deployed revision and retain verification evidence.
Evidence coverage
- SSVC priority
- MEDIUM
- Confidence
- Not available
- Aliases
- None reported
Scores appear only when corresponding evidence exists. Missing data is never interpreted as zero risk.
Sources and record history
Record history
Published Sep 25, 2020 · Last refreshed Nov 3, 2026