Key Takeaways
- A critical sandbox‑escape flaw (CVE‑2026-6875, CVSS 9.5) in the ServiceNow AI Platform enables unauthenticated remote code execution.
- Defused Cyber observed active in‑the‑wild exploitation targeting the pre‑authentication endpoint
/assessment_thanks.dovia HTTP POST. - ServiceNow released patches in June 2026 across multiple release trains (Brazil EA/GA, Australia Patch 2, Zurich Patch 7b/9, Yokohama Patch 12 HF1b/13).
- The vulnerability allows full compromise of the ServiceNow instance and any connected proxy servers.
- Beyond applying patches, ServiceNow is tightening sandbox restrictions to limit executable code in sandbox contexts.
- Organizations running self‑hosted ServiceNow should prioritize immediate patching, verify endpoint hardening, and monitor for anomalous POST requests to
/assessment_thanks.do. - The incident highlights the growing attack surface of AI‑integrated enterprise platforms and the need for rigorous sandbox validation.
Overview of the Vulnerability
CVE‑2026-6875 is a sandbox escape vulnerability discovered in the ServiceNow AI Platform, a component that integrates generative AI capabilities into ServiceNow workflows. The flaw received a CVSS v3.1 score of 9.5, reflecting its potential to allow an unauthenticated attacker to execute arbitrary code with the privileges of the ServiceNow application server. According to Defused Cyber, the vulnerability stems from insufficient validation of user‑supplied data that is passed into a sandboxed execution environment, enabling an attacker to break out of the sandbox and invoke system‑level commands. Because the affected code path is reachable without authentication, any internet‑exposed ServiceNow instance is initially at risk.
Technical Details of the Sandbox Escape
The vulnerable code resides in a preprocessing routine that handles requests to the endpoint /assessment_thanks.do. When a POST request is processed, certain parameters are deserialized and fed into a Java‑based sandbox that is intended to safely evaluate custom scripts used by AI models. Researchers found that by crafting a malicious payload that abuses reflection APIs and class‑loader mechanics, an attacker can instantiate objects outside the sandbox’s permitted namespace. This manipulation leads to a code‑execution primitive equivalent to the one demonstrated in the original proof‑of‑concept (PoC) exploit, albeit via a different gadget chain. The end result is the ability to run arbitrary operating‑system commands under the ServiceNow service account.
Observed Exploitation Activity
Defused Cyber reported that threat actors have begun exploiting CVE‑2026-6875 in the wild shortly after the vulnerability’s public disclosure. The observed attacks consistently target the same pre‑authentication endpoint (/assessment_thanks.do) using HTTP POST requests containing the malicious payload. Although the exact payload varies, the underlying technique leverages the sandbox‑escape gadget identified by Searchlight Cyber. The exploitation pattern indicates automated scanning tools that probe for vulnerable ServiceNow instances, followed by immediate attempts to establish a foothold—such as deploying web shells or exfiltrating configuration data—once code execution is achieved.
Patch Availability and Version Coverage
ServiceNow addressed CVE‑2026-6875 through a series of updates released throughout June 2026. The fixes are present in the following release trains:
- Brazil EA and Brazil GA
- Australia Patch 2
- Zurich Patch 7b and Zurich Patch 9
- Yokohama Patch 12 Hot Fix 1b and Yokohama Patch 13
Organizations running any of these versions should verify that the latest cumulative update has been applied. For releases not listed, ServiceNow advises contacting support to obtain a back‑port or to consider upgrading to a patched train. The patches not only close the sandbox‑escape vector but also include additional hardening measures described below.
Mitigation Recommendations from ServiceNow
Beyond applying the official patches, ServiceNow is enhancing instance security by “severely restricting the type of code that can run in sandbox contexts.” This involves tightening the whitelist of permissible classes and methods within the AI sandbox, disabling dangerous reflection capabilities, and enforcing stricter input validation on data supplied to the /assessment_thanks.do endpoint. Administrators are advised to review sandbox configuration files, ensure that any custom script approvals follow the principle of least privilege, and enable detailed logging for sandbox‑related events to detect abnormal activity promptly.
Broader Implications for AI‑Enabled Enterprise Platforms
The active exploitation of CVE‑2026-6875 underscores a emerging trend: attackers are increasingly targeting the integration points between traditional enterprise software and AI components. Sandboxed environments, while designed to mitigate risk, can become a weak link when complex data‑flow paths involve deserialization, reflection, or dynamic class loading. Enterprises should treat AI‑augmented modules as critical attack surfaces, applying the same rigorous patch management, vulnerability scanning, and penetration testing regimes used for core applications. Additionally, adopting runtime application self‑protection (RASP) tools that monitor sandbox behavior can provide an extra layer of defense against zero‑day techniques similar to the one observed here.
Conclusion and Action Steps
In summary, CVE‑2026-6875 represents a high‑severity sandbox escape in the ServiceNow AI Platform that is already being leveraged by threat actors to achieve unauthenticated remote code execution. The vulnerability’s reach extends to any connected proxy servers, making a full‑instance compromise possible. ServiceNow’s June 2026 patches across multiple release trains effectively remediate the flaw, and the vendor is simultaneously tightening sandbox restrictions to reduce future risk. Organizations should:
- Apply the appropriate patch for their ServiceNow release train immediately.
- Validate that the
/assessment_thanks.doendpoint rejects malformed POST requests (e.g., via web‑application firewall rules). - Review sandbox configurations and enforce least‑privilege script approvals.
- Monitor logs for unexpected sandbox escapes or anomalous POST activity to the affected endpoint.
- Consider upgrading to a newer, supported release train if running an outdated version that cannot be patched.
By taking these steps, enterprises can close the current exploitation window and harden their AI‑enabled ServiceNow environments against similar future threats.

