LangGraph Task Result Cache Weak Hash (CVE-2026-14742)

Technical Analysis

The vulnerability resides in the _freeze function within langgraph/_internal/_cache.py, which computes cache keys using a weak hashing algorithm when the default_cache_key argument is used. This could theoretically allow an attacker to engineer hash collisions, potentially causing incorrect cache hits and returning stale or mismatched task results to a workflow. The entry point is the task caching mechanism itself rather than an external tool or prompt injection vector, and exploitation would require the attacker to influence cache key inputs, which is non-trivial in most deployments. There is no indication this crosses agent or tool trust boundaries directly; it is a data-integrity weakness in an internal caching subsystem rather than an actively exploitable remote code execution or privilege escalation path.

Affected Systems

LangGraph, LangChain

Detection Signatures

• Monitor for anomalous cache hit patterns returning results inconsistent with input task parameters
• Audit logs for use of default_cache_key with attacker-influenced or low-entropy inputs
• Review LangGraph version in dependency manifests (affected versions up to 1.2.4)
• Watch for unexpected reuse of stale task outputs across differing task invocations

Remediation Steps

1. 1

   Upgrade LangGraph

   Monitor the langchain-ai/langgraph repository for the merged fix and upgrade to the patched version once released.

2. 2

   Avoid weak default cache keying

   If configuring custom cache key functions, use a cryptographically strong hash (e.g., SHA-256) instead of relying on default_cache_key behavior.

3. 3

   Isolate cache trust boundaries

   Ensure task result caches are not shared across trust boundaries or user sessions where a collision could leak or misapply another user's cached results.

4. 4

   Validate cache integrity

   Add integrity checks or task-specific salting to cache key derivation to reduce risk of collision-based result substitution.