Cybersecurity daily — 28 October 2026: separate detection, delegation and runtime authorization
Zuma/Agent Gateway runtime authorization and Socure Agent Intelligence: distinguish release stages and authority boundaries.
Il testo dell'articolo non è disponibile in questa lingua; viene mostrato l'inglese.
Window: 27 October 2026 08:00 to 28 October 2026 08:00 JST. Two material items were selected. Saviynt’s publication metadata records 27 October 17:36:44 JST; Socure’s publication date is known but its exact time relative to the cutoff is unverified. Following the prior IAM Helper report, the focus shifts from connecting an agent to authorizing a specific action now. Both items are vendor announcements, not independent safety evaluations.
## 1. Saviynt: authorization at the tool-call boundary
On 27 October Saviynt described runtime authorization between Zuma and Google Cloud Agent Gateway. Zuma makes an allow/deny decision before a tool call and the gateway enforces it. The company also claims intent-deviation and anomaly defenses; its article does not demonstrate general resistance to prompt injection.
Architecture implications: authorize individual operations separately from MCP connection authentication. Do not trust a model's stated purpose alone. Independently check the caller, target data, arguments and delegation scope at execution. Test denial when the decision service fails, direct-API bypass and argument changes after approval. Watch Zuma and Agent Gateway; this is not a newly finalized OAuth standard.
Primary source: [Saviynt, 27 October](https://saviynt.com/blog/google-cloud-agent-gateway-runtime-authorization). Confirm integration availability, failure behavior and measured detection performance before adoption.
## 2. Socure: separate agent detection from delegated credentials
On 27 October Socure announced Agent Intelligence. Agent Detection is generally available, Agent TrustScore is in beta, and Agent Credential is planned for Q1 2027. Identifying agent-like traffic and proving who delegated which authority have different properties and release stages.
Architecture implications: do not turn a detection score into authorization. Separately verify principal relationships, permitted actions, expiry and revocation. Credential adoption still requires attenuation on redelegation and checks at execution. Do not count a planned credential feature as available infrastructure, or equate a proprietary trust score with an interoperable standard implementation.
Primary sources: [Socure blog, 27 October](https://www.socure.com/blog/introducing-agent-intelligence) / [Socure-issued launch release (availability stages)](https://www.businesswire.com/news/home/20261027789869/en/Socure-Launches-Agent-Intelligence-to-Bind-Every-Agent-to-the-Verified-Human-Behind-It). Marketing accuracy figures are omitted because independent evaluation conditions were not established.
## Decisions
Act now: inventory credentials and paths allowing agents to bypass the gateway. Validate: fail-closed decision-service behavior, target/argument binding and delegation revocation. Watch: actual credential availability and verification specifications. This search found no new primary-source development in TEE/isolation that changes architecture decisions.