HighVulnerability

CVE-2026-31253: flash-attention insecure deserialization in checkpoint loading via torch.load

Source: NVD(opens in a new tab)
Identifier: CVE-2026-31253
Published: May 11, 2026
Record updated: October 9, 2026

Affected: flash_attn <= 2.8.3
Fixed in: No fixed version was stated when the source was last read.
Known exploitation: Not listed in the CISA Known Exploited Vulnerabilities catalog at the last check.
EPSS: 0.2%

Summary

CVE-2026-31253 affects the flash-attention training framework through commit e724e2588cbe754beb97cf7c011b5e7e34119e62 (2025-13-04), which contains an insecure deserialization flaw (CWE-502) in its checkpoint loading. The load_checkpoint() function in checkpoint.py and the checkpoint loading code in eval.py call torch.load() without weights_only=True, so arbitrary Python objects can be deserialized via the pickle module. A victim who loads a maliciously crafted checkpoint during model warmstarting or evaluation has arbitrary code executed on their system.

Mitigation

The source does not state a fix yet. Check the original advisory for updates.

Details

Scores and exploitation

EPSS: 0.2%Estimated probability that this CVE is exploited in the next 30 days, from the FIRST Exploit Prediction Scoring System. Read on November 7, 2026.
Disclosure date: May 11, 2026

Weaknesses and techniques

AI threat technique (MITRE ATLAS): AML.T0010

Affected

Packages: flash_attn@<= 2.8.3

Classification

Attack type: Supply chain
Attack sophistication: Trivial
AI component targeted: Framework
Impact: Confidentiality,
Integrity,
Availability
Labels: security

How this record was produced

Summaries, mitigations and labels on this page are generated automatically from the source. The linked source is authoritative. The source gives this advisory a description and no title, so the title above is a headline generated from the source text.

Original source: https://nvd.nist.gov/vuln/detail/CVE-2026-31253

First tracked: May 11, 2026 at 02:10 PM ET, from NVD/CVE Database

Classifier: claude-haiku-5-5, prompt v4

Classifier confidence: 85% (self-assessed)

Affected packages in the Exposure Registry

Matched by package name and ecosystem. Each entry shows whether the package delegates to a language model and how many tracked packages depend on it.

flash-attnPyPINo LLM dependency detected · 6 tracked dependents

Summary

Mitigation

Affected packages in the Exposure Registry

Related items