dbcve.org
Code InjectionWeakness · CWE-94

CVE-2026-31253

HIGH · 7.3 CVSS v3.1 Published 2026-05-11
Mitigation only
No fix yet — a mitigation exists. There is no fixed release. A documented workaround reduces exposure in the meantime.
See remediation →
82/100
Remediation priority · High
Remotely reachable No privileges Zero-click

Official description Straight from the sourceThe vendor's or NVD's own wording, published unedited. Authoritative, but often terse — it says what broke, rarely what to do.

NVD · unedited
The flash-attention training framework thru commit e724e2588cbe754beb97cf7c011b5e7e34119e62 (2025-13-04) contains an insecure deserialization vulnerability (CWE-502) in its checkpoint loading mechanism. The load_checkpoint() function in checkpoint.py and the checkpoint loading code in eval.py use torch.load() without enabling the security-restrictive weights_only=True parameter. This allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by providing a maliciously crafted checkpoint file. When a victim loads this checkpoint during model warmstarting or evaluation, arbitrary code is executed on the victim's system.

Technical summary Written by usOur analysis, written from the advisory, the CVSS vector and the affected-version data. It adds context the advisory leaves out, and never invents facts that are not in the source.

dbcve analysis · high confidence

The flash-attention training framework uses torch.load() in checkpoint.py (load_checkpoint function) and eval.py without the weights_only=True parameter, allowing deserialization of arbitrary Python objects via pickle. Attackers can craft malicious checkpoint files that execute arbitrary code when victims load them for model warmstarting or evaluation.

MitigationAdd weights_only=True to all torch.load() calls in checkpoint.py and eval.py, or implement cryptographic signature verification for checkpoint files before loading to ensure integrity and provenance.

Verify against the referenced sources before acting — the references below are authoritative for this CVE, this summary is not.

CVSS breakdown How the score is builtThe industry scoring standard. It rates how the flaw is reached, what it takes to exploit, and what an attacker gains — the score is derived from those, not the other way round.

From the vector
Attack vector
Network
Complexity
Low
Privileges
None
User interaction
None
Scope
Unchanged
Confidentiality
Low
Integrity
Low
Availability
Low

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:L

Am I affected? How to checkSteps we derive from the advisory and the affected-version data, so you can decide whether this CVE reaches your setup. They are a guide, not a scan — your own configuration is the authority.

dbcve checks

Work through these to decide whether this CVE applies to you.

  1. Locate flash-attention installation
    Run 'pip show flash-attn' or 'pip list | grep -i flash' to find the package installation path
    Affected if The flash-attention package is installed on the system
  2. Find checkpoint.py and eval.py files
    Use 'find /path/to/site-packages -name "checkpoint.py" -o -name "eval.py"' within the flash-attn package directory
    Affected if These files exist in the flash-attention package
  3. Inspect torch.load() calls in checkpoint.py
    Open checkpoint.py and search for 'torch.load(' patterns, then check if weights_only=True parameter is present in each call
    Affected if Any torch.load() call lacks weights_only=True parameter in checkpoint.py
  4. Inspect torch.load() calls in eval.py
    Open eval.py and search for 'torch.load(' patterns, then check if weights_only=True parameter is present in each call
    Affected if Any torch.load() call lacks weights_only=True parameter in eval.py
  5. Verify checkpoint loading is used
    Search for usage of load_checkpoint function or checkpoint loading in training/evaluation workflows to confirm the vulnerable code path is reachable
    Affected if The code loads checkpoint files without weights_only=True protection

A system is affected if flash-attention is installed and any torch.load() call in checkpoint.py or eval.py lacks the weights_only=True parameter, allowing malicious checkpoint files to execute arbitrary code.

Generated from the published advisory. Verify against your own configuration.

Check your environment

Paste your version and any relevant configuration and it will be compared against the affected criteria above. Do not include secrets or credentials.

AI-assisted, checked against the advisory. Informational, not a guarantee.

Remediation Closing itWhat it takes to close this. Where a vendor fix exists we point at it; where none exists we say so plainly, and can build one. Effort estimates are scoped from the advisory, not from your codebase.

dbcve · scoped
Mitigation available No clean upgrade yet — mitigate in the meantime
Mitigation

Add weights_only=True to all torch.load() calls in checkpoint.py and eval.py, or implement cryptographic signature verification for checkpoint files before loading to ensure integrity and provenance.

Recommended fix Moderate confidence
  1. 1. Locate the load_checkpoint() function in checkpoint.py and identify all torch.load() calls.
  2. 2. Locate the checkpoint loading code in eval.py and identify all torch.load() calls.
  3. 3. Add the weights_only=True parameter to each torch.load() call to prevent arbitrary object deserialization.
  4. 4. If the checkpoint loading requires loading custom Python objects (not just tensors), implement a custom unpickler or register safe reducers instead of using weights_only=True.
  5. 5. Test the modified checkpoint loading with your actual checkpoint files to ensure compatibility.
  6. 6. Alternatively, migrate to a secure checkpoint format like SafeTensors which does not use pickle and is not vulnerable to deserialization attacks.
Caveat If your checkpoints contain custom Python objects (non-tensor data), adding weights_only=True may break loading and require refactoring to use SafeTensors format or a custom loader

Generated from the published advisory — verify against the referenced sources before acting.

Have this fixed Scoped from the published advisory
  • Consultation3.0 h
  • Implementation4.0 h
  • Testing6.0 h
  • Review / QA2.0 h
15.0 hours of engineering $2,580
Get help mitigating

An estimate, not a bill — we confirm scope with you before any work starts. Need it this week? Rush from $4,128.

Scan for this in your stack

Free · runs locally
dbcve dependency scanner

Check whether your project pulls in CVE-2026-31253 — or any other known-vulnerable package — straight from your lock files. Free and open source; it runs locally and uploads nothing.

Get it on GitHub Scan in the browser

References Go to the primary sourcePrimary sources — vendor advisories, patches and trackers. Where our summary and a reference disagree, the reference wins.

Primary sources

Practitioner notes

Contributed

Peer-ranked notes from engineers who’ve handled CVE-2026-31253 in production — separate from our analysis above.

No notes yet

Be the first to add a field note for this CVE — a mitigation you’ve verified, a version caveat, or a link to a working fix. Sign in above to contribute.

What this is

A place for practitioners to share what actually worked: a mitigation you’ve tested, a configuration change, a version- or environment-specific caveat, or a link to a verified patch. The most useful notes rise to the top as peers upvote them, so the signal stays high.

What belongs here
  • Verified mitigations, workarounds, and config changes
  • Version or environment caveats, and links to real fixes
  • No weaponised exploit code, or anything meant to cause harm
  • No spam, self-promotion, credentials, or personal data