Threat Model¶

API Server (`rune_bench/api_server.py`)¶

STRIDE Category	Threat	Mitigation
Spoofing	Unauthenticated caller impersonates legitimate client	Token-based auth on all endpoints
Tampering	Attacker modifies job payload in transit	HTTPS/TLS in production; input validation
Repudiation	No audit log of API calls	Structured request logging with timestamps
Information Disclosure	Error responses leak internal paths/stack traces	Sanitise error messages in production mode
Denial of Service	Unbounded request size causes OOM	Request size limits; rate limiting
Elevation of Privilege	API endpoint grants admin operations without authorisation	Role checks on destructive endpoints

STRIDE Category	Threat	Mitigation
Spoofing	Malicious CRD submission by unprivileged user	Kubernetes RBAC restricts CRD writes
Tampering	Controller reconciles tampered CRD spec	Admission webhook validates spec fields
Repudiation	No audit trail for reconciliation actions	Operator logs all reconcile events
Information Disclosure	Operator logs expose sensitive config	Secrets redacted from structured logs
Denial of Service	Flood of CRD creates exhausts controller queue	Work queue depth limits
Elevation of Privilege	Operator service account over-privileged	Least-privilege RBAC role

STRIDE Category	Threat	Mitigation
Spoofing	Malicious process connects to driver socket	Unix socket permissions; process ownership check
Tampering	JSON payload modified before driver receives it	Schema validation at driver boundary
Repudiation	Driver actions not auditable	All driver invocations logged with action + params hash
Information Disclosure	Driver output contains secrets from environment	Output scrubbing for known secret patterns
Denial of Service	Malformed JSON causes driver crash	Strict JSON schema validation; exception handling
Elevation of Privilege	Driver runs with elevated host permissions	Drivers run as non-root; capabilities dropped

STRIDE Category	Threat	Mitigation
Spoofing	Attacker uses stolen Vast.ai API token	Token stored in env var only; never in source
Tampering	Response from Vast.ai API is MITM-tampered	HTTPS with cert validation; response schema validation
Information Disclosure	API key exposed in logs	Token masked in all log output
Denial of Service	Runaway provisioning incurs excessive cost	Fail-closed cost estimation gate (95% confidence floor)

STRIDE Category	Threat	Mitigation
Spoofing	Attacker points OLLAMA_HOST to malicious server	URL normalisation; warn on non-localhost targets
Tampering	Model response manipulated by MITM	HTTPS for remote Ollama; local-only recommended for production
Information Disclosure	Prompt content logged verbatim	Configurable log level; sensitive prompts redacted
Denial of Service	Ollama warmup blocks all other models	Warmup is intentional; document expected behaviour