Four Critical Langflow Bugs, One Bulletin, and a Metasploit Module Already Waiting
IBM's Oct 28 Langflow advisory names four flaws, two of them unauthenticated 9.8s, with 1.13.0 as the only fix and auto-login as the default that gets you there
On October 28, 2026, IBM published a security bulletin covering four new flaws in Langflow, the open-source visual LLM-agent builder it maintains. The Hacker News writeup on November 2 laid out the headline numbers: two of the four CVEs are rated 9.8 and reachable with no account, every release from 1.0.0 through 1.12.8 is affected, and there is no workaround — upgrade or stay vulnerable. A Metasploit module targeting the authentication bypass is already public.
This is the fifth distinct unauthenticated-RCE-tier CVE cohort against Langflow we've covered. The others were CVE-2025-3248 (the pre-auth code endpoint that JADEPUFFER weaponized for an AI-agent-run ransomware operation), CVE-2026-55255 (cross-tenant IDOR added to CISA KEV), CVE-2026-41264 (Python deny-list bypass with its own Metasploit module), CVE-2026-12940 (MCP stdio env-var injection), and CVE-2026-105740 (MCP stdio command injection, 1,423 honeypot attempts in a month). Different endpoints, different layers, same shape every time: the framework accepts something an attacker can shape into code that ends up executing on the host.
What the four flaws are
Per IBM's bulletin and the Hacker News writeup:
- CVE-2026-93680 — 9.8, no account needed. Authentication bypass. The underlying mechanism is Langflow's "Automatic Login" feature, which was on by default in every release up to 1.12.8 and makes any caller a superuser with no prompt. 1.13.0 flips it off by default. The public Metasploit module author rates the vulnerable range narrower — 1.5.0 through 1.11.4 — and the module lists a target Langflow server's tools without running code.
- CVE-2026-110186 — 9.8, no account needed. Arbitrary code execution via improper control of code generation, reachable through MCP tool enumeration. This is the second unauth 9.8 in the same bulletin and the one that converts the auth-bypass primitive into host code execution.
- CVE-2026-110193 — 8.8, low-privileged account. Langflow's Agentic Assistant can be instructed to produce code that runs OS commands on the host. The attack is the assistant's own output being executed.
- CVE-2026-110192 — 6.5, authenticated. Sensitive information disclosure via weakly protected credentials.
Separately, IBM's October 29 records include CVE-2026-104339, rated 9.8. The bulletin lists it but the Hacker News article does not describe it in detail.
Fixed version for all of these: Langflow 1.13.0, released October 22. IBM lists no workaround for operators who cannot upgrade yet. None of the four flaws was in CISA's KEV catalog as of October 30, per the article.
Why the auth-bypass flag matters more than the CVE number
The fix for CVE-2026-93680 isn't just the version bump. The version bump turns a default off. On 1.12.8 and earlier, the Automatic Login feature is on by default and any caller the server sees — authenticated or not — is a superuser. On 1.13.0 that flag is off. The entire delta on the authentication layer is a feature flag.
That matters because:
- The Metasploit module targets the on-by-default state, not the version string. An operator who's upgraded but who kept the flag on, or who re-enabled it to script against the API, is still in range of the public module.
- "Upgraded" is not the same as "safe." Langflow operators who patched to 1.12.4 or 1.12.5 for IBM's earlier bulletins never had this flag flipped. 1.13.0 is the first release where "default install" doesn't accept arbitrary callers as administrators.
- The 9.8s compose. CVE-2026-93680 gets you superuser; CVE-2026-110186 gets you code execution reachable through MCP tool enumeration; CVE-2026-110193 lets a low-privileged user ask the Agentic Assistant for OS-command code. Taken together, one unauthenticated request can land as code on the host, with the agent framework's own code-generation path doing the lift.
What the chain looks like
Set the CVE numbers aside and watch what the Langflow process is being asked to do when CVE-2026-93680 and CVE-2026-110186 compose:
- An unauthenticated HTTP request lands against the server.
- The Automatic Login code path returns a superuser session without a credential check.
- The caller enumerates MCP tools attached to the server.
- One of those enumerations reaches a code-generation path whose output ends up executed in the worker's process context.
No traditional exploit payload is needed. The server renders the attacker's request as a code-generation prompt and runs the output. That is the shape of this whole class of agent-framework CVE: the framework's own code-execution path is the attack vector, and the "payload" is whatever the attacker gets the assistant to write.
What to actually do if you run Langflow
The practical steps are the ones IBM and the article name:
- Upgrade to 1.13.0. It is the only release that fixes all four. There is no workaround.
- If you have upgraded but kept Automatic Login on, turn it off. The Metasploit module targets the on-by-default state; the version string isn't enough.
- Assume the public Metasploit module is being used for discovery scans. Treat any Langflow server that was internet-reachable before the upgrade as suspect.
- If you were running 1.12.8 or earlier on an internet-reachable server at any point since October 22, rotate anything the server had access to — LLM provider keys, cloud credentials, database passwords, anything in its environment or secret store.
- Watch what the Langflow worker actually does. The CVE numbers describe how the attacker gets in; the runtime signal is the server generating code that reaches for shell state the server has no business touching.
Fourth time this year an unauthenticated-RCE-tier CVE cluster has landed on Langflow in a single bulletin or disclosure. The pattern isn't going away by itself.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.