Google-Authored Agent Mesh Shipped With Two Critical Authorization Gaps
Google published two critical authorization CVEs on the same day against its open-source Agent Mesh / SAM project — a cross-machine substrate for exposing MCP tools, Agent-to-Agent traffic, and inference endpoints to a federated mesh of agents. Both carry a CVSS of 9.3. Google is the CNA on both — these are Google disclosing flaws in Google's own alpha code, not an outside find against a GA Google service.
One upfront caveat worth stating plainly: the project's own package page says "This is not an officially supported Google product." SAM is Google-authored, Google-hosted (github.com/google/sam, Apache-2.0), and currently at v0.1.0-alpha.9. Treat it as research-tier infrastructure that happens to live under Google's org, not as Google Cloud.
The two flaws
CVE-2026-103091 sits at the mesh's identity front door. The OIDC endpoints GET /info and POST /register accept unauthenticated requests and enroll any holder of a valid OIDC token with no claim filtering and no ban check. The default policy then binds sam:system:authenticated to wildcard services and targets, so the self-enrolled principal gets read, write, and execute on connected MCP backend nodes. Affected: 0.1.0-alpha.1 through 0.1.0-alpha.8. Fixed in 0.1.0-alpha.9. Credit: Eli Ainhorn of Noma Security — the same researcher who disclosed SneakPeek against Anthropic's Claude Code Desktop five days earlier.
CVE-2026-102503 skips the identity layer altogether. The /libp2p-http ingress server (StartIngressServer in cmd/sam-node/node.go) lacks the Biscuit-token verification and mesh-enrollment checks that the MCP stream handler enforces. An unauthenticated remote libp2p peer can invoke registered backend services directly — including command/stdio-backed MCP servers, A2A, and inference endpoints. Affected: 0.1.0-alpha.1 through 0.1.0-alpha.2. Fixed in 0.1.0-alpha.3.
Neither CVE has a public exploit at publication. Neither is on CISA's KEV list. Neither has been reported active in the wild.
Why two ingresses on the same mesh matter
If you build on a mesh, you depend on its identity tier the same way you depend on your cloud provider's IAM. Here are two separate ingresses on the same mesh, disclosed by the project's own maintainers on the same day, with two different failure shapes:
- The identity ingress accepts anyone and emits credentials.
- The parallel ingress forgets to go through the identity layer entirely.
A mesh is only as strong as its weakest ingress, and this disclosure pair is a very public statement that the identity layer has to be enforced per ingress, not per mesh. The vendor's own patch pulled the two routes in two separate alpha releases.
This is the same architectural shape the industry has been seeing elsewhere all month. Pillar's Oct 2 Unsloth Studio disclosure found Studio's capability-probe load_model_config honored the safety path while FastLanguageModel.from_pretrained did not — the same code reached through different entry points, with different trust. The Oct 14 SneakPeek write-up in Claude Code showed a branch where the CSP and request filter went silent. Per-primitive branching inside the agent's own code is where the trust boundary lives — and the fix has to cover every branch, not just the common-case one.
What a defender actually sees
The two CVEs produce distinct runtime shapes.
For CVE-2026-103091, the observable is an unauthenticated POST /register arriving at the mesh's OIDC endpoint and resulting in a credential issuance, followed by that new principal exercising broad backend tools it did not previously have. The timing between enrollment and first privileged call is short. The principal doesn't appear in prior access logs.
For CVE-2026-102503, the observable is traffic arriving on /libp2p-http that then invokes MCP, A2A, or inference endpoints without having transited the Biscuit-verification path those endpoints normally see. If the mesh emits authorization-trace events, the trace for a legitimate call includes a Biscuit check; the attacker's call doesn't.
Either observable lives outside the model and outside the agent's reasoning loop — it is a runtime pattern on the mesh's ingress and backend. That is where runtime observation on those surfaces belongs.
The takeaway
Shipping an agent mesh means publishing a trust floor, and that floor has to hold at every door the mesh exposes. The disclosure cadence here is the right one — same-day paired CVEs assigned by the maintainers, named discoverers, two alpha releases closing both doors. Anyone experimenting with SAM should check which alpha they're pinned to and audit any registrations issued during the vulnerable window. Everyone else building federated-agent infrastructure: watch the ingresses, not just the identity statements, and remember that an "it's just an alpha" label does not change what the code does when it's running.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.