Global Cybersecurity Risk Briefing

01 / Daily Analysis

Back to the four-story overview ↑

AWS fixes MCP Server script sandbox validation issue

Facts

AWS disclosed CVE-2026-115845 on 23 October 2026. The advisory says AWS MCP Server's run_script tool used an incomplete denylist when validating Python code, potentially allowing an authenticated remote user to execute arbitrary code through a crafted script within their own isolated session. This is an AWS-managed service; the vendor says it has fixed the issue on the service side and customers do not need to apply a patch.

Analysis

Anti-Ransomware analysis: this issue shows why AI agent security boundaries should not depend on a script sandbox alone. Even after a server-side fix, code execution can amplify misuse and the impact of a compromise if an agent also holds privileged credentials or excessive tools. The advisory does not indicate that any customer has been compromised.

Suggested Actions

• Retain MCP tool-call, identity and result records, and confirm that sensitive operations can be traced.
• Apply least privilege to narrow the AWS identities, regions, resources and tools available to agents. Require additional approval for sensitive changes.
• Confirm that you use the AWS-managed MCP Server. For self-hosted deployments, forks or similarly named components, check their versions and sandbox design separately.

Scope: as described by the source

Source

• Amazon Web Services · CVE-2026-115845 - Incomplete validation in the run_script Python sandbox in AWS MCP Server · Source published 2026.10.24 03:00 HKT · Verified 2026.10.26 08:58 HKT

02 / Daily Analysis

Back to the four-story overview ↑

No patched version for AWS Multi Model Server: remaining users should prioritise migration

Facts

On 23 October 2026, AWS disclosed two issues in the unmaintained Multi Model Server (MMS). A crafted model archive could write files outside the intended directory and potentially execute code with the permissions of the MMS process; the model download path could also be directed to unintended network destinations. AWS says all released versions (1.1.11 or earlier) are affected, has no plans to provide a patch, and recommends discontinuing use and migrating to a supported solution.

Analysis

Anti-Ransomware analysis: issues that allow file writes and requests to internal destinations may increase the risk of malware installation, access to internal services and lateral movement. The advisory does not report ransomware exploitation. As the product is no longer maintained, temporary blocking alone cannot eliminate the long-term risk.

Suggested Actions

• Immediately inventory GitHub, PyPI, Docker images and legacy ML inference environments for remaining MMS deployments.
• Before migrating, restrict the management interface to loopback or trusted sources and prohibit externally supplied model download URLs. If the current configuration cannot separate management and inference interfaces, prioritise isolating the service.
• Set a deadline for migration to a supported model-serving solution, and check service account permissions, network egress and verifiable backups.

Scope: as described by the source

Source

• Amazon Web Services · CVE-2026-115812 and CVE-2026-115813: Issue in Amazon Web Services Multi Model Server · Source published 2026.10.24 01:30 HKT · Verified 2026.10.26 08:58 HKT

03 / Daily Analysis

Back to the four-story overview ↑

Sweet Security launches virtual patching for cloud and AI workloads

Facts

Sweet Security launched Virtual Patching on 19 October 2026. According to the vendor, security teams can deploy vulnerability-specific protection policies from the vulnerability catalogue to running cloud and AI workloads, blocking relevant requests until a permanent patch is installed. Policies persist as workloads scale and restart, and can be removed after patching. These performance and blocking capabilities are vendor claims.

Analysis

Anti-Ransomware analysis: virtual patching can shorten the exposure window between vulnerability disclosure and a formal update, especially for workloads that cannot be taken offline immediately. It is a compensating control, however, and should not replace software patches, compatibility testing, least privilege, network segmentation or verifiable backups.

Suggested Actions

• Before adopting it, measure false positives, latency, coverage and bypass risks on test workloads.
• Assign an owner, a deadline for the permanent fix and removal conditions to each virtual patch, so temporary measures do not become permanent.
• Correlate virtual patching alerts with EDR, network monitoring and backup recovery drill records to verify whether the combined controls are effective.

Scope: as described by the source

Source

• Sweet Security · Sweet Security Launches Virtual Patching to Protect Production Before Permanent Fixes Are Deployed · Source published 2026-10-19 · Verified 2026.10.26 08:58 HKT

This article was prepared with assistance from SI agents. Factual summaries, vendor claims and Anti-Ransomware analysis are presented separately. It provides general information and education, not professional advice for a particular environment. Completeness, timeliness, accuracy and product security are not guaranteed, nor are prevention, decryption or recovery. Check the original sources and consult qualified professionals about your environment. Inclusion does not imply certification, endorsement or partnership. Trademarks and rights to original text belong to their respective owners.