🛡️ High-Risk Langflow Vulnerability Still Unpatched, Honeypot Logs 1,423 Real Attacks | AI Security Weekly

Opening

The most unsettling thing this week is that the high-risk vulnerability in Langflow, the AI workflow platform, has already been logged by honeypots for over 1,400 real attack attempts, not theoretical speculation. Meanwhile, Microsoft patched a record-breaking 1,413 vulnerabilities in a single month, and LiteLLM, a popular LLM gateway, was found to have authorization flaws. And how did an Italian bank get scammed out of nearly 100 million euros by an AI-cloned voice? I'll tell you in a moment.

This Week's AI Security Focus

1. Critical Langflow RCE Vulnerability Under Active Attack

What happened: CVE-2026-105740 (CVSS 9.9, a vulnerability severity score out of 10) in the AI workflow platform Langflow allows authenticated users to execute OS commands on the server, a class of flaw commonly known as RCE (remote code execution), affecting all versions prior to 1.9.0. A honeypot set up by security firm Previdian has logged 1,423 attack attempts from 27 different IPs since September 19, peaking at 333 attempts on October 11, with 66 attempts still recorded on October 15.
Scope of impact: Any Langflow deployment still running a version before 1.9.0 with its admin interface exposed to the internet is at risk, especially teams connected to internal systems or sensitive data.
Defense recommendations: Upgrade to 1.9.0 or later immediately, and check whether the admin interface is unnecessarily exposed to the public internet; bring it behind the internal network or restrict access with an allowlist.
Source: The Hacker News (https://thehackernews.com/2026/10/langflows-critical-authenticated-rce.html)

2. Microsoft's October Patch Tuesday Breaks Record with 1,413 Vulnerabilities

What happened: Microsoft's October 2026 Patch Tuesday fixed 1,413 vulnerabilities in a single release (1,102 in Windows, 134 in Office, 61 in developer tools, 36 in SharePoint Server), already exceeding the total for all of 2025, with the 2026 cumulative count surpassing 4,000. SecurityWeek reports that this surge is largely driven by AI-assisted vulnerability discovery, including an actively exploited Windows shell zero-day, CVE-2026-96951 (CVSS 8.8), which can bypass execution restrictions on files downloaded from the internet.
Scope of impact: All Windows and Office users are affected, and IT teams need to reassess whether their patching cadence can keep up with this volume.
Defense recommendations: Prioritize the actively exploited CVE-2026-96951, then triage the rest based on asset exposure and risk level.
Source: SecurityWeek (https://www.securityweek.com/microsoft-patches-record-1413-vulnerabilities-including-exploited-windows-zero-day/)

3. Italian Bank Loses 95 Million Euros to AI Voice Impersonation Scam

What happened: Fideuram, the private banking arm of Italy's Intesa Sanpaolo, was defrauded of 95 million euros (about $108 m i l l i o n) a f t e r s c a m m e r s u s e d A I - c l o n e d v o i c e c o m b i n e d w i t h m e s s a g i n g a p p s t o i m p e r s o n a t e s e n i o r e x e c u t i v e s . S o f a r 53 m i l l i o n e u r o s h a v e b e e n r e c o v e r e d, a n d M i l a n p r o s e c u t o r s a r e i n v e s t i g a t i n g a s u s p e c t r e p o r t e d l y r e s i d i n g o u t s i d e E u r o p e . A c o n c u r r e n t l y r e l e a s e d i d e n t i f A I r e p o r t e s t i m a t e s t h a t g e n e r a t i v e A I c o u l d p u s h U S f r a u d l o s s e s t o$ 40 billion by 2027.
Scope of impact: Any corporate process that relies on phone calls or voice messages for executive authorization could be replicated by the same tactics, with finance and treasury departments at the highest risk.
Defense recommendations: Large wire transfers or executive-instruction workflows should never rely solely on voice or a single messaging channel for identity verification; add an independent secondary confirmation channel.
Source: Biometric Update (https://www.biometricupdate.com/202610/italian-bank-loses-108m-to-ai-messaging-fraud-as-identifai-warns-of-deepfake-threat)

4. LiteLLM Gateway Found with Authorization and SSRF Flaws

What happened: Check Point Research conducted a security assessment of LiteLLM, a widely used LLM (large language model) proxy layer in enterprises, and uncovered CVE-2026-35029 (the /config/update endpoint lacks authorization checks) and CVE-2026-59823 (SSRF via the user_config parameter, a technique that tricks a server into accessing internal network locations it shouldn't reach), along with two credential exposure and insecure hashing issues tracked under GHSA-69x8-hrgq-fjj8. The tested version was v1.77.7; the fix is available in v1.83.9 and later.
Scope of impact: Enterprises using LiteLLM as a multi-model management gateway should check their version, since a breach of this kind of infrastructure effectively exposes control over the entire AI access layer.
Defense recommendations: Upgrade to v1.83.9 or later, and re-examine authorization settings on admin endpoints as well as credential storage practices.
Source: Check Point Research (https://research.checkpoint.com/2026/when-llm-gateways-go-wrong-findings-from-litellm/)

What happened: Scan data from Censys shows that publicly exposed AI-related software grew from 164,088 IPs in October 2025 to 267,915 by July 2026, a 63% increase over nine months across 44 products. The automation platform n8n accounted for 113,926 hosts, 76.1% of which are still running older versions with high-risk vulnerabilities; the AI agent platform OpenClaw saw its exposed panel count surge to 85,478 within just a few months of launch.
Scope of impact: Any team that puts AI automation tools or agent platforms directly on the public internet may be part of these numbers, especially deployments without a habit of continuous upgrades.
Defense recommendations: Audit your organization's internet-facing AI-related services, check whether versions are outdated, and move anything that can be brought behind the internal network or access controls off the public internet.
Source: Help Net Security (https://www.helpnetsecurity.com/2026/10/15/exposed-ai-infrastructure-report/)

6. Underground Forums Using Stolen AI API Keys for Automated Scanning

What happened: Bitsight's threat intelligence team tracked 16 underground forums (February to August 2026) and found a so-called "autonomous penetration testing" platform called MONST3R-AI, which lets buyers paste in stolen OpenAI, Anthropic, and Google API keys and set budget caps, with the scanning costs actually billed to the original key owners. This allows attackers to conduct reconnaissance without running their own models, making them harder to trace.
Scope of impact: Any organization that has ever leaked an AI provider API key may currently be footing the bill for someone else's attacks, often without realizing it.
Defense recommendations: Treat leaked AI provider keys with the same severity as leaked AWS credentials, verify the provenance of MCP (the protocol that lets AI agents connect to external tools and data sources) servers and agent plugins, require human approval before local coding agents access sensitive directories, and watch for anomalous automated login patterns on Notion, Slack, or GitHub.
Source: Help Net Security (https://www.helpnetsecurity.com/2026/10/16/stolen-ai-api-keys-autonomous-scans/)

7. FTC Issues Civil Investigative Demands to OpenAI and Anthropic

What happened: The US Federal Trade Commission (FTC) has issued civil investigative demands (CIDs) exceeding 50 pages each to OpenAI and Anthropic, requesting documentation on security incidents, internal safeguards, marketing materials, and privacy policies, with the investigation focused on possible consumer protection violations. According to Axios, the FTC will subsequently seek testimony from AI executives and plans to issue CIDs to third-party AI organizations as well.
Scope of impact: Enterprises using services from either provider may subsequently be asked to produce their vendor's written documentation on incident response and safeguards.
Defense recommendations: When evaluating or adopting AI provider services, proactively request written documentation on containment, monitoring, access control, and incident response, rather than waiting until something goes wrong.
Source: Axios (https://www.axios.com/2026/10/15/ftc-openai-anthropic-ai-safety-investigation)

Closing

The one thing to remember this week is that when AI-related infrastructure has a problem, what's at stake is usually the entire access chain, not a single feature. From Langflow to LiteLLM, it's the same lesson. Take ten minutes this week to audit which AI tools in your company still have admin interfaces exposed to the internet. See you next Thursday.