Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-103090 Overview
CVE-2026-103090 affects Google Open Source Software Agent Mesh (SAM) versions 0.1.0-alpha.1 through 0.1.0-alpha.8. The flaw combines Missing Authentication for Critical Function [CWE-306] and Insufficiently Protected Credentials [CWE-522] in the control plane enrollment status handler. An unauthenticated remote attacker can retrieve an approved peer's signed Biscuit identity token through the HandleEnrollStatus function at GET /enroll/status. The attacker then replays the token as a bearer credential to read mesh authorization policies and overwrite router multiaddrs.
Critical Impact
Unauthenticated attackers can steal signed peer identity tokens, disclose mesh authorization policies, and hijack router lease addresses to redirect agent traffic.
Affected Products
- Google OSS Agent Mesh (SAM) 0.1.0-alpha.1
- Google OSS Agent Mesh (SAM) versions through 0.1.0-alpha.8
- Control plane component internal/controlplane/server.go
Discovery Timeline
- 2026-10-21 - CVE-2026-103090 published to the National Vulnerability Database
- 2026-10-22 - Last updated in NVD database
Technical Details for CVE-2026-103090
Vulnerability Analysis
The flaw resides in the SAM control plane's enrollment status handler. The HandleEnrollStatus function at GET /enroll/status in internal/controlplane/server.go returns a peer's signed Biscuit identity token without authenticating the caller. An attacker only needs a peer_id value, which the unauthenticated GET /info endpoint exposes directly.
A second flaw compounds the issue. The POST /enroll route lets any holder of a bootstrap token re-fetch another peer's credential. Together these endpoints form a credential-exfiltration primitive against the control plane.
Once the attacker holds a valid Biscuit token, they replay it as a bearer credential on authenticated endpoints. The /policies endpoint discloses mesh authorization policies, revealing trust relationships and access rules. The /routers/lease endpoint accepts arbitrary router multiaddrs, letting the attacker overwrite advertised addresses and redirect peer traffic through attacker-controlled infrastructure.
Root Cause
The control plane treats possession of a peer_id as sufficient proof to release the associated credential. There is no cryptographic challenge binding the request to the enrollment private key. Rate limiting is absent on credential-releasing endpoints, and router lease addresses are not validated before being published to the mesh.
Attack Vector
The attack requires only network access to the control plane. The adversary first calls GET /info to enumerate peer identifiers. They then issue GET /enroll/status?peer_id=<target> to receive a signed Biscuit token. The token is replayed against /policies and /routers/lease to extract authorization state and poison routing metadata.
See the GitHub Security Advisory GHSA-hp3x-79wr-rx66 for the full technical description.
Detect Identity Attacks the Moment Authentication Fails
Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.
Stop Identity AttacksDetection Methods for CVE-2026-103090
Indicators of Compromise
- Unauthenticated requests to GET /enroll/status carrying peer_id values enumerated from GET /info.
- Repeated POST /enroll requests from a single bootstrap-token holder targeting multiple peer identifiers.
- POST /routers/lease requests overwriting existing router multiaddrs with external or unexpected addresses.
- Access to /policies from source addresses that did not complete a prior enrollment handshake.
Detection Strategies
- Baseline the volume and source distribution of control plane enrollment traffic and alert on deviations.
- Correlate peer_id values returned by /info with the originators of subsequent /enroll/status requests.
- Flag Biscuit token reuse across source IPs that do not match the originally enrolled peer.
- Monitor router lease records for sudden multiaddr changes outside of planned deployment windows.
Monitoring Recommendations
- Enable verbose access logging on internal/controlplane/server.go endpoints, including request bodies for /enroll and /routers/lease.
- Forward control plane logs to a central analytics tier that supports anomaly detection on enrollment patterns.
- Retain enrollment and policy-read audit trails for at least 90 days to support post-incident reconstruction.
How to Mitigate CVE-2026-103090
Immediate Actions Required
- Upgrade SAM to version 0.1.0-alpha.9 or later as published in the GitHub Release v0.1.0-alpha.9.
- Rotate all bootstrap enrollment tokens and Biscuit credentials issued by vulnerable control plane instances.
- Audit /policies reads and /routers/lease writes performed since deployment of any affected alpha release.
- Restrict network reachability of the control plane to trusted management segments until the patch is applied.
Patch Information
Version 0.1.0-alpha.9 requires cryptographic proof of possession of the enrollment private key on credential-releasing endpoints. It also enforces rate limiting on enrollment handlers and validates router lease addresses before publication. Full release notes are available in the GitHub Release v0.1.0-alpha.9.
Workarounds
- Place the SAM control plane behind an authenticating reverse proxy that rejects unauthenticated requests to /enroll/status, /enroll, /info, /policies, and /routers/lease.
- Enforce mutual TLS between peers and the control plane to prevent arbitrary network callers from reaching enrollment endpoints.
- Apply upstream rate limiting on enrollment paths to slow credential enumeration until the patched release is deployed.
# Example reverse-proxy allowlist pattern - block unauthenticated enrollment access
location /enroll/ {
allow 10.0.0.0/8;
deny all;
auth_request /_mtls_verify;
proxy_pass http://sam-controlplane;
}
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

