LangChain Framework Hit With Critical CVEs
LangChain and LangGraph have patched three high-severity flaws that could expose files, API keys, and chat histories across AI apps.
Intelligence analysis by GPT-5.4 Mini
Security researchers say three critical vulnerabilities in LangChain and LangGraph could let attackers access files, secrets, and stored conversations. The fixes matter because the frameworks sit deep in the AI stack and are widely reused.
A popular set of AI building blocks had three serious holes, like weak locks on different doors. If attackers find them, they might peek at files, steal secret keys, or read saved chats.
Analysis
What was found
TechJuice reports that LangChain and LangGraph patched three high-severity vulnerabilities affecting widely used AI development frameworks. The flaws are identified as CVE-2026-34070, CVE-2025-68664, and CVE-2025-67644.
The first issue is a path traversal bug in LangChain that can allow arbitrary file access without validation. The second is a critical deserialization flaw that may let attackers exfiltrate API keys and environment secrets through prompt injection. The third is an SQL injection problem in LangGraph’s SQLite checkpoint implementation, which could allow query manipulation.
According to the article, security researcher Vladimir Tokarev said exploitation could expose sensitive files such as Docker configurations, leak secrets through prompt injection, and reveal conversation histories tied to sensitive workflows. That makes the impact broader than a single bug in a single package.
Why it matters
The article emphasizes that LangChain is deeply embedded across the AI ecosystem, with many libraries wrapping or depending on it. That means vulnerabilities can spread through downstream tools and integrations that inherit the same unsafe code paths.
The reported fixes are specific: update langchain-core to 1.2.22 for CVE-2026-34070; update to 0.3.81 or 1.2.5 for CVE-2025-68664; and update langgraph-checkpoint-sqlite to 3.0.1 for CVE-2025-67644.
Security guidance in the piece also warns developers to audit code that passes external or user-controlled configuration into load functions, disable secrets_from_env defaults, and treat LLM output as untrusted input. The core message is straightforward: AI app stacks need the same defensive discipline as any other internet-facing software.
Key points
- LangChain and LangGraph patched three high-severity vulnerabilities.
- The bugs could expose files, API keys, environment secrets, and conversation histories.
- Researchers warn the impact can spread through downstream libraries and integrations.
- The article lists specific fixed versions for each CVE.
- Developers are urged to audit load functions and treat LLM output as untrusted input.
If developers upgrade quickly, the patched versions can close the main exposure paths before attackers find them. The article also gives concrete defensive steps, which could help teams reduce risk beyond this specific issue.
If teams keep older versions in production, the flaws could expose secrets, files, or workflow history. Because LangChain sits inside many other tools, one missed update could spread the risk across multiple AI systems.
