V Vanderburgh.it
News updated

AI News Nuggets

Persistent agents and shared AI platforms need controls around memory, cost, and capacity

Wiz has documented Gemini API abuse billed through hijacked Google Cloud accounts. Databricks has opened agent memory in beta, and Google has released llm-d 1.0 for distributed inference. Together they make three operating boundaries more concrete: who may attach AI spend, what an agent retains, and how a shared serving fleet handles competing workloads.

Editorial read

This edition collects 3 notes across 3 topic areas and 3 sources. Start with Treat cloud billing links and AI API keys as privileged access paths, Give persistent agent memory an owner, scope, and deletion path, Test inference scheduling and isolation with the traffic mix your agents actually create to get the week's main practical signal before scanning the remaining links.

Edition signal

The October 23 signal is to govern the resources AI uses after a prompt ends

AI operations now include persistent state, billing relationships, and reusable model-serving capacity. Wiz's incident research shows how stolen cloud privileges can turn a victim billing account into a funding source for AI running in external projects. Databricks' beta makes agent memory a managed store that needs ownership and deletion rules. Google's llm-d release adds scheduling and cache controls for busy inference fleets. Map and test these boundaries before expanding access or scale.

SecurityAgentsTools
Security
Incident research

Treat cloud billing links and AI API keys as privileged access paths

Source: Wiz CIRT

Wiz's incident-response team reports multiple Google Cloud intrusions in which attackers used stolen privileged credentials, weakened policy controls, and attached their own projects to victims' billing accounts. The attackers then ran Gemini API workloads in external projects, leaving the victim with charges but without visibility into the prompts or outputs. Wiz reports unauthorized charges reaching tens to hundreds of thousands of dollars within 24 to 48 hours in the cases it investigated.

Why this matters: Review who can administer billing accounts, attach projects, change domain restrictions, and create AI API keys. Alert on new external billing links, unfamiliar billing administrators, policy changes, unexpected Gemini API enablement, and sharp spend increases. Test a response path that can revoke access and detach rogue projects quickly; a cost alert alone may arrive after substantial usage.

Read Wiz's cloud billing incident analysis
Agents
Official beta release

Give persistent agent memory an owner, scope, and deletion path

Source: Databricks

Databricks has introduced Agent Memory in beta for Agent Bricks. Managed sessions retain conversation and tool history; a separate memory store keeps selected preferences, domain facts, and workflows; and background pipelines consolidate or replace stale entries. The service runs on Lakebase and Databricks says developers can inspect, export, and delete retained entries. Its reported quality and token gains come from internal benchmarks, not independent production results.

Why this matters: Pilot with a narrow data class and named user scope. Verify which conversations and tool results become long-term memory, whether source permissions still govern retrieval, how corrections propagate, and whether export and deletion reach both sessions and distilled memory. Measure task quality and cost on your own workflow before adopting the beta broadly.

Read Databricks' Agent Memory announcement
Tools
Official project release

Test inference scheduling and isolation with the traffic mix your agents actually create

Source: Google Cloud

Google says llm-d 1.0 is an open-source, Kubernetes-native stack for serving models across distributed accelerator pools. It combines cache-aware and session-aware routing, multi-tenant flow control, tiered key-value cache, and high-availability components. These capabilities target bursty multi-step agent traffic alongside interactive requests; Google's reported performance examples are workload-specific and should not be treated as general results.

Why this matters: Replay representative interactive and background traffic before moving a shared model fleet. Measure tail latency, throughput, cache reuse, queueing, priority enforcement, and upgrade behavior under burst load. Confirm the components your serving engines and accelerators actually support, then set tenant quotas and service targets from the observed results.

Read Google's llm-d 1.0 release