Skip to content
Cloud Security DeskSearch
Menu

Research topic

AI systems

AI risk begins before the model endpoint.

38 publications

Threat modeling for data paths, orchestration identities, model services, and the cloud systems around AI workloads.

View every topic

From the desk

38 publications
Technical guide · October 10, 2026 · Source-based analysis

Run AI coding agents in GitHub Actions without exposing secrets

Treat the agent job as an untrusted-input processor. Follow the outside text from the trigger that admits it to the credentials it can reach, then split reading from writing so a successful injection reaches nothing worth stealing.

Technical guide · October 10, 2026 · Source-based analysis

Treat a cloud model retirement as a security change

Bedrock and Agent Platform fail closed on a model's retirement date, while Foundry can swap a different model in behind the same deployment name. Read the three calendars, compare the replacement's controls and decide per deployment.

Technical guide · October 10, 2026 · Source-based analysis

Admit MCP servers before agents can call their tools

An MCP server's tool descriptions can change after approval with no new release. Record provenance and execution context, hash the definitions per authorization context, enforce the pin at a gateway or allowlist, and watch for drift.

Technical guide · October 10, 2026 · Source-based analysis

Lock down self-hosted model servers before exposing them

Built-in credentials in vLLM, SGLang and Triton cover less than their names suggest, Ollama's local API has none, and vLLM and SGLang leave their internal planes unauthenticated. Here is what stays open and what must sit in front.

Technical guide · October 10, 2026 · Source-based analysis

Validate the judge before trusting automated AI evaluations

Managed agent evaluators score your traces with a second model. Measure that judge against your own reviewers with chance-corrected agreement, probe its biases, and pin the judge model before its pass rate gates a release.

Technical guide · October 10, 2026 · Source-based analysis

Keep semantic response caches inside authorization scope

A semantic cache returns whatever its nearest stored prompt earned. Put tenant, user or entitlement scope, template and model into an exact-match partition, close the write path, and treat thresholds as a correctness setting.

Technical guide · October 8, 2026 · Source-based analysis

Map the OWASP Top 10 for Agentic Applications to cloud controls

Each OWASP agentic risk mapped to the controls AWS, Microsoft and Google document for their agent platforms, with a plain account of which risks no infrastructure setting can close.

Technical guide · October 8, 2026 · Source-based analysis

Design AI agents that contain indirect prompt injection

Filters lower the odds that an agent obeys injected text. Architecture decides what an obeying agent can reach. Compare six published patterns and CaMeL by what each removes, what it costs and what it still misses.

Technical guide · October 8, 2026 · Source-based analysis

Know where cloud AI services process your prompts

Bedrock inference profiles, Foundry deployment types and Agent Platform endpoints each decide where a prompt is processed, separately from where data rests. Compare what each provider commits to and how to enforce it.

Technical guide · October 8, 2026 · Source-based analysis

Compare managed prompt attack filters in Bedrock, Azure and Vertex AI

Bedrock Guardrails, Azure Prompt Shields and Google Cloud Model Armor all score text for injected instructions. They differ in which content reaches the classifier, how you mark it, and what each one skips.

Technical guide · October 8, 2026 · Source-based analysis

Run code written by AI agents inside a disposable sandbox

Treat agent-generated code as untrusted input. Pick a microVM or user-space kernel boundary, keep egress and credentials off by default, and destroy the session when the user or conversation ends.

Technical guide · October 8, 2026 · Source-based analysis

Track EU AI Act deadlines after the Digital Omnibus delay

The Digital Omnibus moved high-risk AI duties to December 2027 and August 2028, but prohibitions, AI literacy and Article 50 transparency already apply. See which dates moved and what cloud teams should build now.

Technical guide · September 12, 2026 · Source-based analysis

Count tokens before sending a prompt to Amazon Bedrock

Count a Bedrock prompt for the chosen model, check endpoint support, distinguish context and output limits, and verify usage without logging prompts.

Technical guide · September 12, 2026 · Source-based analysis

Remove a document from an Amazon Bedrock knowledge base

Retire an S3 document from a Bedrock knowledge base through source changes, sync results, retrieval checks, and review of older application copies.

Technical guide · September 12, 2026 · Source-based analysis

Call Vertex AI from Cloud Run without a service account key

Use a Cloud Run service account and ADC to call Vertex AI, separate deployment and runtime permissions, and remove an obsolete prototype key.

Technical guide · September 12, 2026 · Source-based analysis

Remove unnecessary personal data before sending Vertex AI prompts

Reduce personal data in Vertex AI prompts with field selection, Sensitive Data Protection, deliberate transformations, and safe failure handling.

Technical guide · September 12, 2026 · Source-based analysis

Handle Azure OpenAI rate limits without retry storms

Separate deployment quota, short request bursts and bounded client retry behavior.

Technical guide · September 12, 2026 · Source-based analysis

Connect to Azure OpenAI without an API key

Prove Entra-based inference with a scoped runtime identity before disabling local authentication.

Technical guide · September 2, 2026 · Source-based analysis

Bind confidential GPU inference to a verified key release

Require composite evidence, owner policy and verified recipient binding before confidential inference receives protected key material.

Technical guide · September 2, 2026 · Source-based analysis

Choose who can share an inference prefix cache

Choose the principals allowed to share prefix state, then carry that decision through request routing, offload, transfer and restore.

Research note · September 2, 2026 · Source-based analysis

Embeddings still need a sensitive data boundary

Treat vector access as a derived-data release decision, with separate permissions for retrieval, debugging and bulk export.

Technical guide · September 2, 2026 · Source-based analysis

Give persistent agent memory an admission boundary

Separate conversational candidates from admitted memory, preserve their permitted audience, and make derived summaries repairable.

Technical guide · September 2, 2026 · Source-based analysis

Isolate document parsing before RAG ingestion

Give document parsing a bounded worker, then admit its extracted content separately before embedding or indexing.

Technical guide · September 2, 2026 · Source-based analysis

Decide what happens when an AI guardrail fails

Distinguish denial, incomplete checks and late results, then decide what may cross each protected release point.

Questions answered

  1. What does Cloud Security Desk cover under AI systems?

    Threat modeling for data paths, orchestration identities, model services, and the cloud systems around AI workloads.

    Supporting context

    AI risk begins before the model endpoint.

  2. Why does the AI systems topic matter to cloud security?

    AI risk begins before the model endpoint. This topic applies that principle to practical control evidence and review decisions.

  3. Which control questions belong to AI systems?

    Research belongs here when its central evidence, failure mode, or operational decision falls within the AI systems boundary described on this page.

  4. How many AI systems publications are available?

    This page currently lists 38 publications assigned to AI systems.

  5. How should readers use the AI systems research?

    Start with the publication closest to the control, platform, or evidence gap under review, then follow its findings, limitations, numbered sources, related work, and downloadable material.

  6. Can AI systems overlap other research topics?

    Yes. A publication can appear under several topics when one evidence path crosses multiple control boundaries; each publication page identifies its complete scope.

    Supporting context