Ollama Possible Memory Exhaustion Resource Abuse
Description
Identifies high Ollama runner counts and large model buffer allocations as possible resource pressure indicators. Memory values vary with model size and hardware, and these operational logs do not establish malicious intent; validate against host capacity and workload baselines.
Query · spl
`ollama_server`
("*llama_kv_cache*" OR "*compute buffer*" OR "*llama runner started*" OR "*loaded runners*")
| rex field=_raw "count=(?<runner_count>\d+)"
| rex field=_raw "(?<memory_mb>[\d\.]+)\s+MiB"
| rex field=_raw "started in\s*(?<load_time>[\d\.]+)\s*seconds"
| rex field=_raw "source=(?<code_source>[^\s]+)"
| bin _time span=5m
| stats count as operations
avg(load_time) as avg_load_time
avg(memory_mb) as avg_memory
dc(code_source) as unique_sources
max(load_time) as max_load_time
max(memory_mb) as max_memory
max(runner_count) as peak_loaded_runners
values(code_source) as code_sources
by _time host
| where operations > 10 OR peak_loaded_runners > 5 OR max_memory > 400
| eval avg_memory=round(avg_memory, 2)
| eval max_memory=round(max_memory, 2)
| eval avg_load_time=round(avg_load_time, 2)
| eval max_load_time=round(max_load_time, 2)
| eval severity=case(peak_loaded_runners > 10, "critical", peak_loaded_runners > 5 OR operations > 20, "high", max_memory > 400 OR operations > 10, "medium", 1=1, "low")
| eval attack_type="Resource Pressure / Memory Allocation"
| table _time host operations peak_loaded_runners
unique_sources avg_memory max_memory
avg_load_time max_load_time severity attack_type
| `ollama_possible_memory_exhaustion_resource_abuse_filter`
Implementation guide
Ingest Ollama logs via Splunk TA-ollama add-on by configuring file monitoring inputs pointed to your Ollama server log directories (sourcetype: ollama:server), or enable HTTP Event Collector (HEC) for real-time API telemetry and prompt analytics (sourcetypes: ollama:api, ollama:prompts). CIM compatibility using the Web datamodel for standardized security detections.
Known false positives
- Legitimate high-volume production workloads processing multiple concurrent requests, users loading large language models (7B+ parameters) that naturally require substantial memory allocation, simultaneous multi-model deployments during system scaling, batch processing operations, or initial system startup sequences may generate similar memory allocation patterns during normal operations.
Analyst notes
Known false positives: Legitimate high-volume production workloads processing multiple concurrent requests, users loading large language models (7B+ parameters) that naturally require substantial memory allocation, simultaneous multi-model deployments during system scaling, batch processing operations, or initial system startup sequences may generate similar memory allocation patterns during normal operations.