SecureAEye LAB ONLINE · SIG 657 · DET ~41s

← All attacks / AI/ML and emerging attacks

MLflow prompt model-version source arbitrary file read (CVE-2026-116690)

An incomplete fix for an earlier MLflow path check let schemeless relative paths be stored as a prompt model-version source, so the artifact endpoint served arbitrary files readable by the server process

categoryAI/ML and emerging attacksseverityHighlabnot yet measured

Example: mlflow (pip) 3.10.0 to before 3.17.0, CVSS 7.5; fixed in 3.17.0; follows incomplete fix of CVE-2026-2614

Classification

34 values: 34 machine-classified. What do these mean?

Exploitability

Attack vectorSoftware vulnerabilities
Reach (CVSS)Network
Attack complexityLow
Attack requirementsPresent
Privileges requiredNone
User interactionNone
Attacker skill requiredLow
Resources requiredLow
Network requiredYes
Likelihood / prevalenceMedium
MaturityEmerging

Impact & scope

SeverityHigh
ImpactConfidentiality
PlatformWeb AppAI SystemLinux
EnvironmentOn-premisesCloudAI system
Target asset classML model / agentServerData store
Kill-chain phase (tactic)Initial AccessCollection
Remote supportYes

Detection & telemetry

Noise levelMedium
Data sources (ATT&CK data components)Application Log ContentFile AccessNetwork Traffic Content
Log sourceswebserver:accessapplication:logsai:model-serving
Defense bypassedSystem access controls
Attack permissions (ATT&CK)User

Detection & response

Mitigated by 2

Framework references

Evidence & sources