How the attack works
vLLM's guided-decoding feature (Outlines backend) compiles a grammar for each unique schema it sees and writes it to a local cache with no cleanup. An attacker scripts a high volume of short requests, each carrying a schema that has never been used before, within a short time window. The cache grows without bound as it stores a new grammar file per request. Eventually the filesystem fills up, causing failures for all tenants sharing that vLLM instance. The detection ties this pattern to a single session by tracing process lineage, so it flags one session's request flood, not cross-session or cross-user correlation.
Why it matters
A single malicious session can exhaust disk space on a shared vLLM server, causing denial of service for every other user or application relying on that endpoint.
What you can do
- Set a size cap and eviction policy on the grammar/schema cache directory, or move it to a bounded, auto-cleaned filesystem.
- Rate-limit or throttle structured-output requests per session/API key, especially bursts of distinct schemas.
- Monitor disk usage on vLLM hosts and alert before it reaches capacity.
- Review high-volume structured-output workloads (benchmarks, fuzzers, CI suites) and separate them from production infrastructure to avoid false alarms and real exhaustion alike.
Known benign look-alikes
- Evaluation or benchmark harnesses that emit a distinct JSON schema per test item (for example per-document field extraction), producing many unique schemas quickly.
- Schema fuzzing or conformance suites run against a self-hosted vLLM endpoint in CI.
- Agent frameworks that inline request-specific enum values, IDs or few-shot examples inside the schema, making every schema textually unique for an otherwise benign workload.
- Deliberate load or capacity testing of a structured-output endpoint.
- Mitigations - the rule requires at least 25 schema-bearing POSTs from ONE process session inside a 5-minute window, a >=90% unique-to-total schema ratio, applies a 10-minute per-session cooldown so a sustained job reports once rather than continuously, and only ever reports (never blocks) so no inference request is interrupted.