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 · August 28, 2026 · Source-based analysis

Security evidence for AI release decisions

A release approval should identify the changed application, the claims its tests support and the evidence that expires when a model, prompt, data path or runtime changes.

Research note · August 28, 2026 · Source-based analysis

What coding benchmarks can prove about a model

A coding benchmark result depends on its tasks, harness and tests. A reproducible count of SWE-bench Verified shows why the denominator belongs beside every comparison.

Technical guide · August 28, 2026 · Source-based analysis

Where fine tuning data needs a trust boundary

A training dataset can preserve its checksum and still teach the wrong behavior. Admission controls need to separate origin, transformation, approved use and the model change they produce.

Technical guide · August 28, 2026 · Source-based analysis

A budget model for bounded AI inference

Request throttles, token quotas and billing alerts control different things. An inference service needs an admission decision that reserves bounded work and reconciles what actually ran.

Technical guide · August 28, 2026 · Source-based analysis

An evidence plan for AI incident response

An AI incident record needs to connect the system version, context, output, downstream action and observed consequence. A saved prompt or a public report is only part of that evidence.

Technical guide · August 28, 2026 · Source-based analysis

Model output needs its own trust boundary

A model can produce valid JSON containing an unauthorized identifier, an unsafe link or text that a renderer interprets as code. The application consuming that output owns the next trust decision.

Technical guide · Published August 28, 2026 · Source-based analysisSeries August 18, 2026

Map every connection beyond the private AI endpoint

Trace inference, retrieval, tools, administration, and telemetry separately before describing an AI application as private.

Technical guide · Published August 28, 2026 · Source-based analysisSeries July 9, 2026

Keep sensitive prompts out of routine telemetry

Collect diagnostic metadata for a defined purpose and treat prompts, retrieved documents, tool arguments, and responses as a separate content-capture decision.

Technical guide · Published August 28, 2026 · Source-based analysisSeries May 30, 2026

Put enforceable boundaries around agent tool calls

Treat model proposals as requests for authority, then check the operation, resource, recipient, and approval at the point where a tool can create a side effect.

Technical guide · Published August 28, 2026 · Source-based analysisSeries April 20, 2026

Treat model downloads as software supply chain changes

Review weights, custom code, dependencies, and runtime authority as separate decisions before promoting a downloaded model into a trusted environment.

Technical guide · Published August 28, 2026 · Source-based analysisSeries March 11, 2026

Keep MCP tokens bound to the intended resource

Keep token audiences, user consent, and downstream tool authority separate when reviewing a protected HTTP MCP service.

Technical guide · Published August 28, 2026 · Source-based analysisSeries January 30, 2026

Keep private documents out of shared RAG answers

Authorize retrieved documents before they enter model context, preserve permissions on chunks, and make source access changes visible in the retrieval path.

Research report · August 26, 2026 · Desk publication

Qwen3.8-Flash-Next and GLM-5.3-Flash share a 3:1 long-context pattern

Both models replace most conventional attention layers with recurrent state and reserve sparse attention for periodic retrieval. Their differences lie in where they place capacity, how much neural computation they activate, and what their serving stacks must keep trustworthy.

Research note · July 14, 2026 · Illustrative analysis

Threat-model the system around the model

The model endpoint is one component. The consequential paths often run through retrieval stores, orchestration identities, evaluation data, and operator tools.