T14, Infrastructure & Economic Warfare

Attacks aimed at degrading service availability, exhausting budget, or causing economic harm via abuse of the LLM endpoint. Adjacent to classical DoS but LLM-specific cost dynamics.

Techniques

T14.001, Per-account budget exhaustion

• Multiple low-bandwidth requests each maximizing token cost
• Free-tier abuse via account creation farms
• Direct DDoS of paid endpoint (uses budget at attack rate)

Defenders: per-account caps + alerts.

T14.002, Shared-quota poisoning

Multi-tenant LLM services share quotas at a provider level. Attacker exhausts the shared quota → all tenants degraded.

Targets: SaaS products that meter LLM access but share underlying API account.

T14.003, GPU resource starvation

Self-hosted inference. Long-context queries hog GPU:

• Send queries near max context length
• Send batch of long-context queries in parallel
• Loop to maintain pressure

Cost-amplification ratio: 1 user request → 100% GPU utilization.

T14.004, Cost amplification (T5.002 cross-ref)

Prompt engineering to maximize output cost:

• "Repeat 'token' 5000 times"
• "Generate the maximum-length response you're allowed to produce"
• Quadratic context patterns that bloat each turn

T14.005, Pricing-tier exploit

Some APIs charge differently based on model variant. Trick the endpoint into using the expensive variant via parameter manipulation (T5.004 overlap).

T14.006, Caching-pollution

Where LLM provider caches identical prompts (cost-reduction feature):

• Submit prompts that fill cache but produce uncacheable outputs
• Cache thrashing → no cost savings, full cost incurred

T14.007, Streaming attack

Open many streaming requests → never consume → endpoint holds connection slots open until timeout.

Probe pattern