29MAY

AI orchestration flaw joins CISA's KEV

4 min read

14:17UTC

CISA added a CVSS 9.4 Langflow flaw and a Trend Micro Apex One bug to its exploited-vulnerabilities catalogue on 21 May, with a 4 June federal patch deadline.

← Back to GitHub's own code cloned via VS Code add-on Jump to analysis ↓

TechnologyDeveloping

Key takeaway

Two AI orchestration flaws reached CISA's exploited catalogue in three weeks, making the control plane a confirmed attack class.

CISA added two flaws to its KEV catalogue of vulnerabilities confirmed exploited in the wild on Thursday 21 May, with a federal patch deadline of 4 June ¹. The first is CVE-2025-34291 in Langflow, an open-source visual builder for stitching together large-language-model agent pipelines, rated CVSS 9.4: an origin-validation error that combines permissive cross-origin resource sharing, missing cross-site request forgery protection, and an endpoint that runs code by design. The second is CVE-2026-34926 in Trend Micro Apex One on-premises, a directory-traversal flaw rated CVSS 6.7.

Langflow stores API tokens and credentials for every downstream software-as-a-service it integrates, which is what makes the flaw dangerous. A single origin-validation bug does not stop at Langflow; it converts into lateral access across every connected service, the same blast-radius logic that made identity providers high-value targets. A security team that has never inventoried its shadow LLM deployments, often stood up by data Teams rather than security, cannot rotate credentials it does not know are exposed.

The Iran-nexus group MuddyWater was already documented abusing this flaw in a March 2026 analysis. The new fact is the 21 May KEV listing, which confirms in-the-wild exploitation and puts a federal clock on it, not that MuddyWater has only just arrived. Paired with the LiteLLM proxy flaw that UNC6780 exploited within 36 hours of its own KEV listing a fortnight earlier , the pattern is no longer theoretical. The AI orchestration layer, the control plane that connects models to data and tools, now carries two KEV entries in three weeks.

Deep Analysis

In plain English

Langflow is an open-source tool that lets technical teams assemble AI systems without writing code from scratch. Organisations use it to build automated pipelines that connect large AI language models to databases, external websites, and internal services. Because it connects to so many other systems, it typically holds the passwords and access keys for all of them in one place. A serious security flaw in Langflow, rated 9.4 out of 10 by the international vulnerability-scoring standard, has been confirmed as actively exploited in the wild. An Iranian-linked hacking group called MuddyWater was documented exploiting it in March 2026. On 21 May 2026, the US government's cyber agency CISA formally added the flaw to its list of vulnerabilities requiring urgent action, with a federal deadline of 4 June. The risk is that a single compromised Langflow instance can hand attackers the keys to every other system it connects to. Many Langflow deployments were set up by data science teams rather than IT security teams, which means they are often not tracked in organisations' standard security monitoring.

Deep Analysis

Root Causes

AI orchestration platforms aggregate credentials for every downstream service they connect: cloud providers, database systems, external APIs, and model endpoints. Langflow's architecture stores these credentials in an internal SQLite or PostgreSQL database and passes them to agent pipelines at runtime.

A successful exploitation of CVE-2025-34291 gives the attacker full access to this credential store, beyond the Langflow instance itself. The blast radius is proportional to the number and privilege level of downstream integrations, not the Langflow deployment's own network position.

The CORS and CSRF design flaws that compose CVE-2025-34291 are a consequence of Langflow's origin as a single-developer prototyping tool that was later deployed in production environments without a security architecture review. The codebase was designed for local development on localhost, where cross-origin and cross-site request forgery protections are conventionally relaxed. When the tool migrated to shared-server deployments, the permissive defaults followed.

MuddyWater's selection of Langflow as a target reflects a documented Iranian threat-intelligence priority: AI platforms that connect to defence and government research networks. Langflow's open-source positioning and its active user community in universities and government data labs make it a predictable target for an actor seeking access to AI-assisted research pipelines rather than financial data.

What could happen next?

Risk
Shadow AI infrastructure, including Langflow instances deployed by data teams without IT security oversight, now represents a confirmed attack surface that can yield access to the full credential set of an organisation's AI-connected services in a single exploitation.
Immediate · Assessed
Precedent
Two AI orchestration tool KEV entries in three weeks (LiteLLM on 8 May, Langflow on 21 May) establishes the AI control plane as a recognised attack class in CISA's mandatory-action framework, which will accelerate enterprise security teams' inclusion of AI tooling in standard vulnerability management programmes.
Short term · Assessed
Consequence
MuddyWater's documented March 2026 exploitation of Langflow, combined with the KEV classification, creates a notification obligation for UK and EU public-sector organisations under NIS2 and the CS&R Bill's 24-hour incident reporting requirement if they were running unpatched instances during the exploitation window.
Immediate · Assessed

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: The 2017 Apache Struts CVE-2017-5638 exploitation that enabled the Equifax breach followed the same structural path: a widely deployed middleware component running in a category (Java enterprise frameworks) that many organisations did not track in their patch management inventories because it was considered developer tooling rather than production infrastructure.

The flaw was added to a US-CERT alert within days of public disclosure; Equifax had the alert but did not correlate it to an unmanaged Struts instance in its network, which had been in continuous exploitation for 76 days before detection. The Langflow case maps onto this structurally: the flaw class is different (origin-validation vs RCE) but the asset-inventory failure is identical.

Counter-parallel: The 2021 Log4Shell (CVE-2021-44228) response demonstrated that when an AI orchestration-equivalent component (a near-universal Java logging library) faces a CVSS 10.0 flaw with simple exploitation, the patch velocity across the ecosystem is faster than for any previous library-level CVE because the vendor community and CERT network mobilised simultaneously within 72 hours.

Langflow's CVSS 9.4 has not triggered equivalent ecosystem mobilisation, partly because of the asset-inventory gap and partly because the MuddyWater exploitation has remained geographically bounded in its confirmed victim set.

Consensus view: Wiz Research's 2025 AI security survey found that approximately 65% of enterprise AI infrastructure stood up by data science and machine-learning Teams had not been included in the organisation's formal software asset inventory, meaning it was invisible to both patch management and network-segmentation controls.

Langflow installations are predominantly deployed outside standard IT change management because they are treated as data-team tooling rather than production application infrastructure. This inventory gap is the mechanism that converts a CVSS 9.4 flaw into an undetected, unpatched network Node with access to every downstream API key the instance holds.

Counter-view: Unit 42 (Palo Alto Networks) researchers argue in their 2026 AI Threat Intelligence report that the KEV classification of AI orchestration tools overstates operational severity for most enterprises because Langflow is rarely exposed directly to the internet in mature deployments; the more common attack path is lateral movement from a compromised internal host to a Langflow instance on the internal network.

On this reading, MuddyWater's March 2026 documented exploitation targeted organisations where data Teams had bound Langflow to a public IP without a reverse-proxy layer, a deployment pattern common in startup environments and academic research networks.

Key tension: Whether the AI orchestration control plane should be classified as critical infrastructure under CISA's KEV framework, triggering mandatory federal patch deadlines, or whether the two AI-tier KEV entries in three weeks (LiteLLM CVE-2026-42208 on 8 May, Langflow CVE-2025-34291 on 21 May) reflect confirmed exploitation of a genuinely high-value target class rather than a policy classification decision.

Sources:The Hacker News·CISA

Mentions:Known Exploited Vulnerabilities →Palo Alto Networks →Unit 42 →Iran →UNC6780 →LiteLLM CVE-2026-42208 →LiteLLM →CVE-2026-20182 →MuddyWater →CVE-2025-34291 →Langflow →Trend Micro Apex One →CISA →

First Reported In

Update #5 · GitHub's own code cloned via VS Code add-on

The Hacker News· 29 May 2026

Read original →

Causes and effects

Caused by

LiteLLM SQL injection hits in 36 hours

Langflow's KEV listing follows LiteLLM being exploited within 36 hours of its own KEV addition, establishing a pattern of the AI orchestration layer as a confirmed active attack class.

Occurred 8 May 2026

Read story →

This Event

AI orchestration flaw joins CISA's KEV

Two AI-tier entries in three weeks move the layer that wires models to data and tools from theoretical risk to a confirmed attack class.

Led to

Langflow hits KEV a second time

Langflow's first KEV entry, exploited by Iran-nexus MuddyWater, is escalated by the second entry two months later, making the framework a recurring exploitation surface.

Occurred 21 Jul 2026

Read story →

Different Perspectives

Group-IB

Group-IB, headquartered in Singapore, documented on 19 March that The Gentlemen began inside Qilin's own affiliate programme and split from it over a payment dispute. Its analysis makes this window's leak-site tally readable as a fracture inside one ransomware operation rather than the arrival of an unknown crew.

ENISA

ENISA placed EU drinking water and wastewater in its cyber risk zone for the first time on 28 May, finding one in three water entities had never run a risk assessment. CISA's 30 July advisory to American operators reads as the operational sequel to a gap European regulators had already documented two months earlier.

House of Lords and NCSC

The Cyber Security and Resilience Bill sits at House of Lords committee stage, its amendment paper running to 23 July. Separately, the NCSC guidance CISA cited when directing water operators to disconnect controllers is version 1.0 from March 2024, a two-year-old British document doing operational work in an American alert published last week.

heise online

heise online reported on 27 July that Russian state attackers, named as APT28, ran the hotel WiFi credential-harvesting campaign. Its attribution is firmer than ReliaQuest's own hedged low-to-medium confidence assessment of the same intrusion, a difference in national reporting posture rather than a settled dispute over who is responsible.

CISA and ReliaQuest

CISA compressed remediation deadlines under BOD 26-04 while, in the same window, telling water and wastewater operators to disconnect internet-exposed controllers rather than patch them. ReliaQuest, investigating the hotel WiFi campaign, held its assessment of the intrusion route at low-to-medium confidence and declined to name a state sponsor.

AIVD

AIVD co-signed AA26-204A alongside CISA, the NSA and the FBI, adding its signature to the fifteen-agency coalition naming LAUNDRY BEAR. Dutch involvement continues a pattern of joint Five Eyes-adjacent advisories the agency has co-issued with CISA and NCSC through 2026.