Breaking

CVE-2026-0768 in Langflow Abused to Harvest OpenAI API Keys and AWS Secrets

September 1, 2026 at 15:59 EDT

Security
AI Agents
Open Source

Threat actors are exploiting a critical unauthenticated remote code execution flaw in Langflow, CVE-2026-0768, to steal OpenAI API keys, AWS secrets, and Langflow superuser credentials from internet-exposed instances. VulnCheck recorded the campaign on United Kingdom canaries, with at least 50 exploitation attempts over the weekend of August 30, 2026, rising to 360 observed attacks as of Monday, September 1.

September 1, 2026 · Langflow

Critical Langflow RCE used to steal cloud keys

Unauthenticated code execution in the validate endpoint is being used to harvest OpenAI API keys, AWS secrets, and Langflow superuser credentials from exposed instances.

9.8

CVSS score (critical)

50+

Weekend canary hits

360

Attempts by Monday

Added Langflow CVEs in 2026

Attempts jumped more than 7× in days

Canary detections: weekend baseline versus Monday total. Each block ≈ 50 attempts.

50+

Weekend

360

By Monday

From unauthenticated request to stolen keys

Hit validate
Unauth POST with a crafted code parameter

→

Exec as root
Python runs without a login or sandbox

→

Dump secrets
Env vars, secret_key, SSH, bash history

→

Steal keys
OpenAI API keys and AWS credentials

What defenders have

Version 1.10.1 removes the unsafe exec() path. Operators are told to move to the latest release, 1.11.6, restrict public access, and rotate secrets.

What still cuts

No public proof-of-concept was required. Patch notes did not clearly name the CVE. Internet-facing hosts still concentrate live OpenAI and AWS keys.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.