AWS discloses a flaw that gave any network client full admin control of Loom, its open-source platform for running AI agents

AWS published security bulletin 2026-124-AWS at 12:00 PM PDT on 2 October. It covers three flaws in Loom for AWS. In the bulletin's words: “Loom is an AWS Labs open-source AI agent orchestration platform.”

The most serious, CVE-2026-103956, affected Loom versions before 1.6.1 that had no identity provider configured. In those, unauthenticated network clients could get full administrative control of the agent control plane. AWS fixed it in version 1.6.1, released on 4 August 2026, and is disclosing it now.

The other two need a signed-in user with certain write permissions. CVE-2026-103957 let such a user make the backend send OAuth2 secrets or access tokens to endpoints an attacker controls. CVE-2026-103958 let such a user send requests to internal network locations, including credential endpoints. Both are fixed in version 1.7.0.

AWS advises upgrading to 1.7.0, rotating OAuth2 secrets and access tokens, and reviewing CloudTrail logs if credentials may have been exposed. The bulletin gives no severity score and says nothing about exploitation. It credits Kenneth Cox. Severity scores shown on third-party trackers are not repeated here.

Agent orchestration platforms hold the keys the agents use. A control plane left open to the network hands an attacker every credential and permission the agents have, which is why AWS's advice goes beyond the patch to rotating secrets.

• Confirmed Bulletin 2026-124-AWS, published 2 October 2026, covers three CVEs in Loom for AWS, an AWS Labs open-source AI agent orchestration platform. AWS
• Confirmed Before 1.6.1, with no identity provider configured, unauthenticated network clients could get full admin control of the agent control plane; fixed in 1.6.1 on 4 August. AWS
• Confirmed Two credential-exposure and internal-request flaws are fixed in 1.7.0; AWS advises upgrading and rotating OAuth2 secrets and tokens. AWS