Proposal: define use cases for AI-relevant SARIF generation and consumption.

[michaelcfanning]

SARIF for AI Security Findings — Scenario Guide

Version: 0.1 (draft)
Status: Proposal

---

Use Cases

Dimension	S1: AI Primary Scanner #775	S2: AI Triages Single Tool] #776	S3: AI Synthesizes Multiple Tools #777
Upstream SARIF	None	One run	Multiple runs
AI's results	New results[] in its own run	Rewrites existing results; originals preserved in original.* property bag	New results[] from synthesis
Tool identity	AI is driver	Original tool untouched; AI provenance in run.properties	All originals preserved; AI has its own run
Rule IDs	AI's own (CWE-based)	Upstream tool's	AI's own (CWE-based)
Taxonomy (taxa)	Rarely needed	Optional — AI may add CWE refs to upstream results	Recommended — maps multiple tool rule namespaces to shared taxonomy
Source cross-references	N/A	suppressions[], level rewrites, ai.triage.confidence	relatedLocations + ai.synthesis.sourceResults
runs[] count	1	1	3+

---

Scenario 1: AI Primary Scanner

AI scans source code directly with no upstream tool output. AI is run.tool.driver. The LLM model and named skills are recorded in run.tool.extensions[]. Findings are first-class result objects with CWE-based rule IDs (CWE-78/mcp-server). CWE-78 covers OS Command Injection across any context; CWE-78/mcp-server is an instance of it in MCP server tool handlers.

#775

---

Scenario 2: AI Triages Single Tool Output

One upstream tool (CodeQL, Semgrep, a custom scanner) has produced SARIF. AI evaluates each finding and issues a verdict: false positive, true positive, or reprioritization. The original tool's identity and results are preserved. AI verdicts are expressed as a lossless overlay: result.suppressions[] for FPs, result.level rewrites with original.level preservation for reprioritization, and ai.triage.* property bag fields for machine-readable triage metadata supporting automated re-triage pipelines.

#776

---

Scenario 3: AI Synthesizes Multiple Tool Outputs

Multiple upstream tools have produced SARIF. AI reads across all runs, correlates patterns that no single tool may have detected alone, and issues its own findings in a new run. Upstream runs MAY be embedded in the same runs[] array when high-fidelity cross-referencing is needed — this enables both structured ai.synthesis.sourceResults property bag references and spec-defined sarif: URI embedded links (§3.10.3) in message text that point directly to upstream result objects by JSON Pointer. These URIs are fragile to post-processing. When upstream runs are absent, contributing code locations are referenced via relatedLocations[] and upstream results by result.guid if available. A taxonomies entry in the AI run maps across upstream tool rule namespaces to a shared CWE taxonomy. AI may also create taxa that map individual upstream tool rules to CWE entries, enabling cross-tool grouping for downstream consumers.

#777

[michaelcfanning]

As I trust is happening elsewhere, Microsoft is exploring changes required to integrate AI into SARIF generating and consuming systems. This work item is intended to be a proposal to define SARIF uses cases that warrant guidance/design thinking. I will create some draft content for the three scenarios proposed here then update links to point to those proposals.

This work item exists to support discussion/argument around the use cases themselves, do these make sense? Can use cases be collapsed? Do we require additional? etc.