Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-108858 Overview
CVE-2026-108858 is a sensitive information exposure vulnerability affecting Predibase LoRAX through version 0.12.1. The router component writes the caller-supplied api_token from POST /generate request bodies into router logs through an instrumented GenerateParameters span field. Any party with access to the router logs or an OpenTelemetry Protocol (OTLP) trace backend can recover other users' private-adapter tokens. The flaw is categorized as a logging-related information disclosure issue tracked under [CWE-532: Insertion of Sensitive Information into Log File].
Critical Impact
Attackers with read access to router logs or OTLP trace storage can harvest private-adapter API tokens belonging to other LoRAX users, enabling impersonation and unauthorized use of fine-tuned model adapters.
Affected Products
- Predibase LoRAX versions up to and including 0.12.1
- LoRAX router component (router/src/lib.rs and router/src/server.rs)
- Deployments exporting router traces to OTLP-compatible backends
Discovery Timeline
- 2026-10-11 - CVE-2026-108858 published to the National Vulnerability Database (NVD)
- 2026-10-16 - Last updated in NVD database
Technical Details for CVE-2026-108858
Vulnerability Analysis
LoRAX is an inference server for serving fine-tuned Low-Rank Adaptation (LoRA) adapters. Clients submit generation requests to the POST /generate endpoint and may supply an api_token field in the JSON body to authenticate access to private adapters hosted on external model registries.
The LoRAX router instruments its request handling with tracing spans. The GenerateParameters struct is annotated such that incoming field values, including api_token, are recorded as span attributes. When the router emits structured logs or exports spans through OTLP, the raw token value is written in plaintext alongside the request. Operators collecting telemetry for performance analysis therefore accumulate a store of long-lived user credentials in log aggregation and tracing systems.
Root Cause
The root cause is the absence of a redaction or #[serde(skip)]-equivalent annotation on the sensitive api_token field inside GenerateParameters. The field is included in the tracing span the same way as non-sensitive generation parameters. Review the referenced router source in lib.rs and the server handler in server.rs for the instrumented code paths.
Attack Vector
Exploitation does not require interacting with the vulnerable endpoint directly. An attacker needs read access to one of the downstream telemetry destinations: router stdout logs, container log collectors, SIEM indexes, or OTLP trace backends such as Jaeger, Tempo, or a vendor-hosted observability platform. Once accessed, the attacker searches recorded spans for the api_token attribute and extracts tokens issued by other tenants. Harvested tokens can be replayed against private adapters to download proprietary models or incur billing against the token owner.
For additional analysis, see the VulnCheck advisory on Lorax token exposure and the HackMD write-up for CVE-2026-108858.
Detection Methods for CVE-2026-108858
Indicators of Compromise
- Log lines or trace spans from the LoRAX router containing an api_token attribute with a non-empty value.
- GenerateParameters spans in OTLP backends where the token field is populated rather than redacted.
- Unexpected downloads of private adapters associated with accounts whose tokens never appeared in direct API traffic.
Detection Strategies
- Grep historical router logs and OTLP span exports for the string api_token to identify previously recorded credentials.
- Audit access logs on log aggregation platforms (Elasticsearch, Loki, Splunk) and tracing backends for queries that search for the api_token field.
- Correlate adapter download events on upstream model registries with the source IPs that generated the original POST /generate request to spot token reuse by a different client.
Monitoring Recommendations
- Instrument alerts on log pipelines when fields matching known secret patterns, including api_token, are ingested from LoRAX routers.
- Monitor telemetry backend authentication events and limit retention of historical traces that contain the vulnerable field.
- Track outbound requests from the LoRAX router to model registries and flag mismatches between the requesting user and the token owner.
How to Mitigate CVE-2026-108858
Immediate Actions Required
- Rotate every api_token that has been submitted to a LoRAX 0.12.1 or earlier deployment, treating all previously used tokens as exposed.
- Purge router logs and OTLP trace stores that contain the api_token attribute, including backups and cold storage snapshots.
- Restrict read access to router logs and tracing backends to the minimum set of operators required for troubleshooting.
Patch Information
At the time of publication, no fixed release is referenced in the NVD entry. Monitor the Predibase LoRAX GitHub repository for a release that redacts the api_token field from the GenerateParameters tracing span. Until an official patch is available, apply the mitigations below.
Workarounds
- Disable tracing export from the router, or set the tracing level so that GenerateParameters spans are not emitted to OTLP collectors.
- Deploy a reverse proxy in front of LoRAX that strips the api_token field from request bodies before they reach the instrumented handler, substituting a server-side credential lookup.
- Apply a local patch that marks the api_token field with a redaction attribute or omits it from the span, following the structure referenced in the HackMD mitigation guide.
# Configuration example: disable OTLP export and reduce router log verbosity
unset OTEL_EXPORTER_OTLP_ENDPOINT
export OTEL_TRACES_EXPORTER=none
export RUST_LOG=lorax_router=warn
# Scrub historical logs containing api_token values
grep -rl 'api_token' /var/log/lorax/ | xargs sed -i 's/"api_token":"[^"]*"/"api_token":"[REDACTED]"/g'
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

