CVE-2026-28449: OpenClaw webhook replay triggers

CISO Take

OpenClaw AI agent framework (pip) fails to maintain durable replay state for Nextcloud Talk webhook events, meaning any attacker who captures a single valid signed HTTP request can replay it indefinitely without suppression. The flaw requires zero privileges, no user interaction, and exploits over the network with low complexity — placing it in the top 82nd EPSS percentile despite a medium CVSS of 6.5. In agentic deployments, each replayed event re-triggers downstream tool invocations, workflow steps, or API calls, compounding both integrity violations and service degradation. CISA has not added this to KEV and no public exploit exists, but the trivial attack bar means any adversary with passive network visibility can weaponize it. Upgrade to OpenClaw 2026.2.25 or apply patch commit d512163; as an interim control, enforce webhook idempotency at the API gateway using request-ID deduplication or short-lived nonces.

Sources: NVD EPSS GitHub Advisory vulncheck.com ATLAS

What is the risk?

Medium priority with elevated operational risk in agentic deployments. CVSS 6.5 understates real-world impact because the CVSS vector (AV:N/AC:L/PR:N/UI:N) reflects a near-zero exploitation barrier. EPSS top 82nd percentile signals active attacker interest relative to peers. The package carries 396 historical CVEs, suggesting a chronic security debt. With only 4 known downstream dependents, blast radius is contained today, but OpenClaw's positioning as an AI agent hub (evidenced by AIID #1368 ecosystem abuse) means unpatched instances may be systematically targeted. No KEV listing, no public scanner template, and no Metasploit module keep this below critical threshold.

How does the attack unfold?

Capture

Adversary intercepts a valid signed POST request to OpenClaw's Nextcloud Talk webhook endpoint via passive network monitoring or a MITM position.

AML.T0006

Replay

Adversary retransmits the captured request one or more times; OpenClaw accepts each replay as legitimate because no durable nonce or timestamp state is checked.

AML.T0091

Agent Tool Invocation

Each accepted replay triggers the AI agent to re-execute its configured tool chain (email dispatch, API calls, data writes) as if the original event recurred.

AML.T0053

Impact

Duplicate agent actions corrupt downstream state (integrity) or exhaust service quotas and processing capacity (availability), degrading the AI agent deployment.

AML.T0029

Capture

Adversary intercepts a valid signed POST request to OpenClaw's Nextcloud Talk webhook endpoint via passive network monitoring or a MITM position.

AML.T0006

Replay

Adversary retransmits the captured request one or more times; OpenClaw accepts each replay as legitimate because no durable nonce or timestamp state is checked.

AML.T0091

Agent Tool Invocation

Each accepted replay triggers the AI agent to re-execute its configured tool chain (email dispatch, API calls, data writes) as if the original event recurred.

AML.T0053

Impact

Duplicate agent actions corrupt downstream state (integrity) or exhaust service quotas and processing capacity (availability), degrading the AI agent deployment.

AML.T0029

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
OpenClaw	pip	—	No patch
4 dependents 37% patched ~3d to patch Full package profile →

Do you use OpenClaw? You're affected.

How severe is it?

CVSS 3.1

6.5 / 10

EPSS

0.3%

chance of exploitation in 30 days

Higher than 19% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What is the attack surface?

AV Network

AC Low

PR None

UI None

S Unchanged

C None

I Low

A Low

What should I do?

5 steps

Patch: Upgrade OpenClaw to version 2026.2.25 or later (patch commit d512163d686ad6741783e7119ddb3437f493dbbc adds durable replay suppression).
Interim workaround: Implement request-ID deduplication at your API gateway or reverse proxy — reject any webhook POST whose X-Request-ID or payload hash has been seen within a 5-minute sliding window.
Add timestamp validation: reject webhook requests with an iat/timestamp older than 60 seconds, forcing attackers to replay within a tight window.
Detection: Alert on duplicate webhook delivery attempts with identical signatures in your WAF or SIEM; monitor for anomalous agent action rates (sudden spikes in tool invocations per user per hour).
Scope reduction: Restrict webhook endpoint exposure to Nextcloud Talk source IP ranges if the deployment topology allows it.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable Yes

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Auth Bypass DoS Agent Plugin AML.T0029 - Denial of AI Service AML.T0034.002 - Agentic Resource Consumption AML.T0049 - Exploit Public-Facing Application AML.T0053 - AI Agent Tool Invocation AML.T0091 - Use Alternate Authentication Material

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.9.2 - Information security for AI systems

NIST AI RMF

MANAGE-2.2 - Processes for AI risk response and recovery

OWASP LLM Top 10

LLM07 - Insecure Plugin Design

Frequently Asked Questions

What is CVE-2026-28449?

OpenClaw AI agent framework (pip) fails to maintain durable replay state for Nextcloud Talk webhook events, meaning any attacker who captures a single valid signed HTTP request can replay it indefinitely without suppression. The flaw requires zero privileges, no user interaction, and exploits over the network with low complexity — placing it in the top 82nd EPSS percentile despite a medium CVSS of 6.5. In agentic deployments, each replayed event re-triggers downstream tool invocations, workflow steps, or API calls, compounding both integrity violations and service degradation. CISA has not added this to KEV and no public exploit exists, but the trivial attack bar means any adversary with passive network visibility can weaponize it. Upgrade to OpenClaw 2026.2.25 or apply patch commit d512163; as an interim control, enforce webhook idempotency at the API gateway using request-ID deduplication or short-lived nonces.

Is CVE-2026-28449 actively exploited?

No confirmed active exploitation of CVE-2026-28449 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-28449?

1. Patch: Upgrade OpenClaw to version 2026.2.25 or later (patch commit d512163d686ad6741783e7119ddb3437f493dbbc adds durable replay suppression). 2. Interim workaround: Implement request-ID deduplication at your API gateway or reverse proxy — reject any webhook POST whose X-Request-ID or payload hash has been seen within a 5-minute sliding window. 3. Add timestamp validation: reject webhook requests with an iat/timestamp older than 60 seconds, forcing attackers to replay within a tight window. 4. Detection: Alert on duplicate webhook delivery attempts with identical signatures in your WAF or SIEM; monitor for anomalous agent action rates (sudden spikes in tool invocations per user per hour). 5. Scope reduction: Restrict webhook endpoint exposure to Nextcloud Talk source IP ranges if the deployment topology allows it.

What systems are affected by CVE-2026-28449?

This vulnerability affects the following AI/ML architecture patterns: AI agent frameworks, Event-driven agent pipelines, Webhook-integrated agent systems, Collaborative AI assistants (chat-integrated agents).

What is the CVSS score for CVE-2026-28449?

CVE-2026-28449 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.27%.

What is the AI security impact?

Affected AI Architectures

AI agent frameworksEvent-driven agent pipelinesWebhook-integrated agent systemsCollaborative AI assistants (chat-integrated agents)

MITRE ATLAS Techniques

AML.T0029 Denial of AI Service

AML.T0034.002 Agentic Resource Consumption

AML.T0049 Exploit Public-Facing Application

AML.T0053 AI Agent Tool Invocation

AML.T0091 Use Alternate Authentication Material

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.9.2

NIST AI RMF: MANAGE-2.2

OWASP LLM Top 10: LLM07

What are the technical details?

Original Advisory

OpenClaw versions prior to 2026.2.25 lack durable replay state for Nextcloud Talk webhook events, allowing valid signed webhook requests to be replayed without suppression. Attackers can capture and replay previously valid signed webhook requests to trigger duplicate inbound message processing and cause integrity or availability issues.

Exploitation Scenario

An adversary with passive visibility into traffic between the Nextcloud Talk server and an OpenClaw agent instance (e.g., via a shared network segment, a compromised upstream router, or a misconfigured load balancer logging full request bodies) captures a POST to the OpenClaw webhook endpoint. The request carries a valid HMAC signature tied to a message event — for example, a user instruction like 'summarize the Q2 financials and email the CFO.' The attacker replays this request 50 times in rapid succession. OpenClaw, lacking replay suppression state, processes each replay as a fresh event: the agent calls its email tool 50 times and floods the CFO's inbox, or saturates a downstream summarization API, triggering a quota breach. If the agent has write access to shared data stores, each replay also writes a duplicate record, corrupting downstream analytics or audit logs.

Weaknesses (CWE)

CWE-294 Authentication Bypass by Capture-replay

CWE-294 — Authentication Bypass by Capture-replay: A capture-replay flaw exists when the design of the product makes it possible for a malicious user to sniff network traffic and bypass authentication by replaying it to the server in question to the same effect as the original message (or with minor changes).

[Architecture and Design] Utilize some sequence or time stamping functionality along with a checksum which takes this into account in order to ensure that messages can be parsed only once.
[Architecture and Design] Since any attacker who can listen to traffic can see sequence numbers, it is necessary to sign messages with some kind of cryptography to ensure that sequence numbers are not simply doctored along with content.

Source: MITRE CWE corpus.