Disclosed weakness · no CVE recorded

Langroid expands model-produced segment ranges into unbounded allocations

Langroid · Langroid relevance extraction

First documented disclosure · Review checkpoint

Summary

An oversized segment range can exhaust memory or discard an otherwise valid retrieval result.

Severity and exploitation evidence

No CVSS score recorded in this review.

Exploitation not established: The researcher explicitly reports no production deployment. Controlled parser tests and maintainer validation establish resource exhaustion, not an observed attack.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The extraction helper materializes every integer in a model-produced range before matching the segments actually present in the passage.

Conditions required

A DocChatAgent relevance extractor emits an oversized range, potentially influenced by retrieved content. The vulnerable extraction path processes it.

Potential impact

The controlled tests consumed gigabytes of memory or raised MemoryError. The caller converts the failure into NO_ANSWER, losing the selected content; unrestricted processes may exhaust available memory.

Affected versions and configurations

Reproduced on 0.69.18 and 0.69.19; the full introduction range is unverified.

Fix and mitigation status

0.69.20 uses interval membership instead of expanding every integer on this extraction path.

Practical next steps

Upgrade to a release containing 0.69.20 and impose memory limits on model-output processing.

Operating lesson

Bound work by actual input data rather than by numeric ranges emitted by a model.

Evidence and limits

The public issue and PR were opened November 5; 0.69.20 shipped later that day. The generic parse_number_range_list API retains its documented expansion behavior. No CVE or CVSS score was verified.

Sources