AWS and GitLab have fixed four security flaws in infrastructure used to manage AI agents. Three affected AWS Loom, a platform for building and operating agents. The fourth affected GitLab’s Self-Hosted AI Gateway.[1][7]

The flaws opened different routes, but all sat in software with power beyond generating text. These control systems decide which tools an agent can use, handle credentials, connect to internal services and run workflows. A weakness there can expose the surrounding organization’s systems.

AWS Loom: administration, secrets and internal services

AWS disclosed three vulnerabilities in Loom.

The first affected Loom versions before 1.6.1. If an organization put Loom’s central server on a network without setting up a login system, Loom could treat anyone who connected as an administrator. An attacker could then add tool servers, read saved credentials and change the cloud permissions given to AI agents. AWS fixed the problem in version 1.6.1 and recommended upgrading to 1.7.0.[1][2]

The second flaw affected connections to tools or remote agents. Before version 1.7.0, a user with permission to manage those connections could feed Loom false login information. Loom could then send a secret or another user’s access token to an outside destination.[1][3][5]

The third flaw also required permission to manage connections. A malicious destination could make Loom contact a private service and return its response to the user. That route could reach services that supply credentials to cloud containers. AWS fixed the flaw in version 1.7.0 and told operators to restrict connection-management permissions while upgrading.[1][4][6]

Together, the three flaws reached several powerful functions gathered in one platform: administrator access, stored secrets, cloud permissions and connections to private network services.

GitLab: a workflow became a server command

GitLab disclosed a separate flaw in some versions of its Self-Hosted AI Gateway. A logged-in user with access to the Duo Agent Platform could submit a malicious workflow, break out of a restricted template environment and run commands on the gateway server.[7][8]

The entry point was an agent workflow. The result was control at the level of the computer running the gateway.

GitLab released patched versions 19.2.4, 19.3.2 and 19.4.1. The company said its hosted gateways had already been fixed.[7]

Why the control layer carries so much power

Agent platforms often bring identity, tools, credentials, network connections and execution into one place so an AI system can complete multi-step work. The AWS and GitLab disclosures show the security consequence of that convenience: a conventional software bug can land at a junction with unusually broad authority.

AWS told Loom operators to upgrade to version 1.7.0. GitLab told operators of affected self-hosted AI Gateways to install version 19.2.4, 19.3.2 or 19.4.1.[1][7]

The code around an agent decides where it may connect, which secrets it may use and what it may execute. These four fixes protect that control layer.

Sources

  1. AWS security bulletin 2026-124-AWS — https://aws.amazon.com/security/security-bulletins/2026-124-aws
  2. CVE-2026-103956 — https://cveawg.mitre.org/api/cve/CVE-2026-103956
  3. CVE-2026-103957 — https://cveawg.mitre.org/api/cve/CVE-2026-103957
  4. CVE-2026-103958 — https://cveawg.mitre.org/api/cve/CVE-2026-103958
  5. AWS Loom advisory GHSA-jcxf-gpf4-58hm — https://github.com/awslabs/loom/security/advisories/GHSA-jcxf-gpf4-58hm
  6. AWS Loom advisory GHSA-w6g6-h8pv-6mc7 — https://github.com/awslabs/loom/security/advisories/GHSA-w6g6-h8pv-6mc7
  7. GitLab AI Gateway 19.4.1 patch release — https://docs.gitlab.com/releases/patches/other-patches/patch-release-gitlab-ai-gateway-19-4-1-released
  8. CVE-2026-90970 — https://cveawg.mitre.org/api/cve/CVE-2026-90970

Methodology · Corrections · Archived original