Lesson 317·M26 AI Security
Model Abuse
OpenLearning Objectives
By the end of this lesson you should be able to:
- Explain the abuse types: the scraping, the cloning, the quota burning (L317)
- Explain the mechanism: the model's access used against you (L317)
- Explain the controls: the rate limits, the auth, the monitoring (L317)
- Explain the cost: the abuse's bill (L334)
- Explain the AI shape: the model's abuse and the controls (L317)
One-Line Definition
The model abuse is the misuse of the model access (L317) — the abuse types (the scraping: the outputs harvested to train a competitor L365; the cloning: the service replicated through its API; the quota burning: the junk calls exhausting the budget L334, L317), the mechanism (the model's access used against you, L317), and the controls (the rate limits L318, the auth L319, and the monitoring L274, L317) — the abuse (L317) stopped at the door (L318).
The one-sentence interview answer: "The model abuse is the misuse of the model access (L317). The types (L317): the scraping (L317) — the outputs harvested (L317) to train the competitor's model (L365) — the service's value (L317) extracted (L317); the cloning (L317) — the service replicated (L317) through its own API (L317) — the competitor (L317) learns the prompts and the flows (L317); and the quota burning (L317) — the junk calls (L317) exhausting the budget (L334) — the legitimate users (L317) starved (L317). The mechanism (L317): the model's access (L278) — the endpoint (L267) with the keys (L319) — used against you (L317): the wide-open endpoint (L317) and the shared keys (L319) are the enablers (L317). The controls (L317): the rate limits (L318) — the per-key (L318) and the per-IP (L318) caps (L317); the auth (L319) — the keys (L319) and the scopes (L319) at the door (L317); and the monitoring (L274) — the anomaly (L317) — the quota spike (L334) and the traffic pattern (L317) — detected (L317). The AI shape (L173): the model endpoint (L278) — the abuse (L317) assumed (L317) and the controls (L317) at the door (L318): the rate limits (L318), the auth (L319), and the monitoring (L274)."
Mental Model
Think of the model abuse as the all-you-can-eat buffet with the freeloaders. The buffet (the model service, L278) serves the paying guests (the legitimate users, L317). The freeloaders (the abusers, L317): the recipe-copiers (the scrapers, L317) — photographing every dish (the outputs, L317) to open their own buffet (the clone, L365); the impersonators (the cloners, L317) — copying the whole experience (L317) through the buffet's own line (the API, L317); and the plate-wasters (the quota burners, L317) — loading the plates and leaving them (the junk calls, L317) — the kitchen (the model, L278) exhausted (L334) and the paying guests (L317) starved (L317). The manager (the controls, L317): the bouncer (the auth, L319) at the door, the per-guest limits (the rate limits, L318), and the cameras (the monitoring, L274) watching the waste (L317). The buffet works because the door is guarded, the limits are set, and the waste is watched (L317).
the buffet (the model service, L278)
┌────────────────────────────────────────────────────────┐
│ the freeloaders (the abusers, L317) — the copiers │
│ (L317), the impersonators (L317), the wasters (L317) │
│ the bouncer (the auth, L319) · the limits (the rate │
│ limits, L318) · the cameras (the monitoring, L274) │
└────────────────────────────────────────────────────────┘
The mental model is the buffet: the freeloaders, the bouncer, the limits, and the cameras (L317).
Visual Flow — One Abuse, Stopped
the abuser (L317)
│ the junk calls (L317) · the scraping (L317)
▼
┌────────────────────── THE DOOR (L317) ─────────────────────────────┐
│ the auth (L319): the key (L319) → the scope (L319) │
│ the rate limit (L318): the per-key cap (L318) → the 429 (L317) │
└──────────────────────────┬───────────────────────────────────────┘
▼
┌────────────────────── THE WATCH (L317) ────────────────────────────┐
│ the monitoring (L274): the quota spike (L334) │
│ the traffic pattern (L317) → the anomaly (L317) │
└──────────────────────────┬───────────────────────────────────────┘
▼
┌────────────────────── THE ACTION (L317) ───────────────────────────┐
│ the key revoked (L319) · the IP blocked (L317) │
│ the alert (L274) │
└──────────────────────────────────────────────────────────────────┘
The flow is the stop: door → watch → action (L317).
How It Works — The Abuse, Part by Part
- The scraping (L317). The outputs harvested (L317) — to train the competitor's model (L365) — the service's value (L317) extracted (L317).
- The cloning (L317). The service replicated (L317) through its own API (L317) — the competitor (L317) learns the prompts and the flows (L317).
- The quota burning (L317). The junk calls (L317) exhausting the budget (L334) — the legitimate users (L317) starved (L317).
- The controls (L317). The rate limits (L318), the auth (L319), and the monitoring (L274) — the abuse (L317) stopped at the door (L318).
Note
The abuse is the access's misuse (L317). The senior answer secures the access (L317): the endpoint (L267) is the door (L317) — the auth (L319) identifies who (L317), the rate limits (L318) bound how much (L317), and the monitoring (L274) watches the pattern (L317). The wide-open endpoint (L317) and the shared keys (L319) are the enablers (L317) — the controls (L317) close them (L317).
Real Project Usage
- A production AI SaaS (L357). The endpoint (L267) — the auth (L319), the rate limits (L318), and the monitoring (L274).
- A model API (L278). The keys (L319) per customer (L320) — the quota (L149) and the rate limits (L318) per key (L317).
- A chatbot product (L162). The session limits (L318) — the junk calls (L317) capped (L317).
- A regulated workload (L371). The abuse (L317) recorded (L322) — the audit (L322) of the access (L317).
- Anything with a model (L317). The abuse (L317) assumed (L317) — the controls (L317) at the door (L318).
The through-line: the controls are the access's guard — the auth, the limits, and the watch (L317).
Interview Explanation
Say it in four moves:
- The scraping. "The outputs harvested to train the competitor (L317)."
- The cloning. "The service replicated through its API (L317)."
- The quota burning. "The junk calls exhausting the budget (L317)."
- The controls. "The rate limits (L318), the auth (L319), the monitoring (L274)."
Senior-Level Insights
- The endpoint is the door (L317). The auth (L319) and the rate limits (L318) at the door (L317) — the abuse (L317) stopped before the model (L278).
- The key is the identity (L319). The per-customer keys (L320) — the quota (L149) and the limits (L318) per key (L317) — the abuse (L317) attributed (L334).
- The quota is the budget (L149). The per-customer quota (L149) — the burning (L317) bounded (L317).
- The anomaly is the watch (L317). The quota spike (L334) and the traffic pattern (L317) — the monitoring (L274) detects (L317).
- The audit is the record (L322). The abuse (L317) recorded (L322) — the attribution (L334) and the legal (L317).
Common Mistakes
- The wide-open endpoint (L317). The model (L278) without the auth (L319) — the abuse (L317) free (L317).
- The shared keys (L319). The one key for everyone (L319) — the abuse (L317) un-attributable (L334).
- The no limits (L318). The unlimited calls (L318) — the quota (L334) burned (L317).
- The watch missing (L274). The spike (L334) undetected (L317) — the abuse (L317) discovered at the bill (L334).
- The audit absent (L322). The abuse (L317) unrecorded (L322) — the attribution (L334) impossible (L317).
Best Practices
- Authenticate at the door (L319) — the per-customer keys (L320).
- Limit per key (L318) — the rate limits (L318) and the quota (L149).
- Watch the anomalies (L317) — the quota spikes (L334) and the patterns (L274).
- Act on the abuse (L317) — the key revoked (L319), the IP blocked (L317).
- Audit the access (L322) — the record (L322) of the who and the what (L317).
Interview Questions
Q: Walk me through the model abuse.
A: The misuse of the model access (L317). The types — the scraping (L317), the cloning (L317), the quota burning (L317). The mechanism — the endpoint (L267) used against you (L317). And the controls — the rate limits (L318), the auth (L319), the monitoring (L274).
Q: What's the scraping?
A: The output harvesting (L317): the abuser (L317) calls the API (L267) at scale (L317) and collects the outputs (L317) — to train the competitor's model (L365) or to replicate the service (L317). The value (L317) extracted through the service's own endpoint (L317).
Q: How do you stop the quota burning?
A: The limits (L317): the rate limits (L318) per key (L318) — the calls per minute (L318) capped (L317); the quota (L149) per customer (L320) — the monthly budget (L334) bounded (L317); and the anomaly detection (L317) — the spike (L334) flagged (L317) and the key revoked (L319).
Q: How do you attribute the abuse?
A: The keys (L319): the per-customer keys (L320) — the calls attributed (L334) to the key (L317). The audit (L322) records the who and the what (L317) — and the bill (L334) shows the abuse (L317).
Follow-Up Questions
- What are the abuse types (L317)?
- What's the scraping (L317)?
- How do you stop the burning (L317)?
- How do you attribute the abuse (L319)?
- What's the watch (L274)?
Comparison Table — The Abuse Types
| The scraping (L317) | The cloning (L317) | The burning (L317) | |
|---|---|---|---|
| The goal (L317) | the data (L317) | the replication (L317) | the denial (L317) |
| The asset (L317) | the outputs (L317) | the flows (L317) | the quota (L334) |
| The control (L317) | the limits (L318), the watermarks (L317) | the auth (L319), the prompts' secrecy (L317) | the quota (L149), the anomaly (L317) |
The senior read: each abuse has its control — the door, the limits, and the watch (L317).
Code Example — The Controls, Applied
// The abuse controls (L317) — the door, the limits, the watch (L317).
// 1 · THE AUTH (L319) — the per-customer key (L320).
const key = await auth.verify(apiKey); // L319
if (!key) return error(401);
// 2 · THE RATE LIMITS (L318) — the per-key caps (L317).
const limited = await rateLimit.check(key.id, {
callsPerMinute: 60, // L318
tokensPerDay: 1_000_000, // the quota (L149)
});
if (!limited.ok) return error(429, 'rate limit'); // L318
// 3 · THE QUOTA (L149) — the budget (L334).
const quota = await quota.check(key.customerId, { monthly: 500 }); // L334
if (!quota.ok) return error(429, 'quota exceeded');
// 4 · THE WATCH (L274) — the anomaly detection (L317).
const anomaly = await monitor.track({
customerId: key.customerId,
tokens: usage.total, // L332
pattern: 'burst', // L317
});
if (anomaly.flagged) await revoke(key.id); // the action (L319)
// 5 · THE AUDIT (L322) — the access recorded (L317).
await audit.log({ key: key.id, action: 'invoke', tokens, at }); // L322
What the reader must SEE — the controls, applied:
auth.verify → the door (L319)
callsPerMinute + tokens → the rate limits (L318)
quota monthly → the budget (L149, L334)
monitor.track + revoke → the watch and the action (L274, L317)
audit.log → the record (L322)
The door guarded, the limits set, the watch watching (L317).
Tip
The pair that defines the controls: the per-key rate limit (the cap, L318) and the anomaly-triggered revoke (the action, L317). Guard the door, cap the calls, watch the pattern, revoke the abuser — the abuse, stopped (L317).
Performance Notes
- The checks are the latency's cost (L317). The auth (L319) and the limits (L318) — the sub-millisecond (L317) at the door (L317).
- The quota is the billing's record (L334). The tokens (L332) per key (L317) — the attribution (L334) at the bill (L317).
- The watch is the pipeline's cost (L317). The anomaly detection (L274) — the metrics (L331) for the pattern (L317).
- The abuse is the bill's spike (L334). The junk calls (L317) — the cost (L334) — the controls (L317) bound it (L317).
Debugging Scenarios
| Symptom | First check (L317) | The lever |
|---|---|---|
| The quota is gone | The burning (L317) | The rate limits (L318), the quota (L149) |
| The outputs are cloned | The scraping (L317) | The watermarks (L317), the limits (L318) |
| The spike is invisible | The watch (L274) | The anomaly detection (L317) |
| The abuse is un-attributable | The keys (L319) | The per-customer keys (L320) |
| The bill exploded | The quota (L334) | The per-customer budget (L149) |
Quick Revision Notes
- The model abuse = the access's misuse (L317): the scraping, the cloning, the burning, the controls.
- The scraping: the outputs harvested (L317) — the competitor's training (L365).
- The cloning: the service replicated through its API (L317).
- The burning: the junk calls exhausting the quota (L334).
- The controls: the rate limits (L318), the auth (L319), the monitoring (L274).
Cheat Sheet
MODEL ABUSE = the scraping, the cloning, the quota burning
THE ABUSE TYPES (L317)
the scraping (L317) — the outputs harvested (L317)
to train the competitor (L365)
the cloning (L317) — the service replicated (L317)
through its own API (L317)
the quota burning (L317) — the junk calls (L317)
exhausting the budget (L334)
THE MECHANISM (L317)
the endpoint (L267) used against you (L317)
the wide-open endpoint (L317) · the shared keys (L319)
THE CONTROLS (L317)
the auth (L319) — the per-customer keys (L320)
the rate limits (L318) — the per-key caps (L318)
the quota (L149) — the per-customer budget (L334)
the monitoring (L274) — the anomaly (L317) → the revoke (L319)
the audit (L322) — the record (L322)
INTERVIEW, 4 MOVES
1 scraping "the outputs harvested (L317)"
2 cloning "the service replicated (L317)"
3 burning "the junk calls, the quota gone (L317)"
4 controls "the auth, the limits, the watch (L317)"Key Takeaways
Recap
- The model abuse is the misuse of the model access (L317): the abuse types (L317), the mechanism (L317), and the controls (L317)
- The abuse types (L317): the scraping (L317) — the outputs harvested (L317) to train the competitor's model (L365); the cloning (L317) — the service replicated (L317) through its own API (L317); and the quota burning (L317) — the junk calls (L317) exhausting the budget (L334)
- The mechanism (L317): the model's access (L278) — the endpoint (L267) with the keys (L319) — used against you (L317)
- The controls (L317): the rate limits (L318) — the per-key caps (L318); the auth (L319) — the per-customer keys (L320); and the monitoring (L274) — the anomaly (L317) detected (L317) and the key revoked (L319)
- The attribution (L317): the per-customer keys (L320) — the calls attributed (L334) and the audit (L322) recording the who and the what (L317)
- The AI shape (L317): the model endpoint (L278) — the abuse (L317) assumed (L317) and the controls (L317) at the door (L318): the rate limits (L318), the auth (L319), and the monitoring (L274)
Check your understanding
Answer each one out loud. Then rate yourself honestly.
0/8What are the abuse types (L317)?
What's the scraping (L317)?
How do you stop the burning (L317)?
How do you attribute the abuse (L319)?
What's the watch (L274)?
What's the quota (L149)?
What's the audit (L322)?
What is the access's misuse (L317)?
A Closing Note — The Buffet, Guarded
You now hold the abuse: the scraping, the cloning, and the burning — with the bouncer at the door and the cameras watching. The freeloaders are capped — and the paying guests are served (L317).
Next: the control that stops the abuse at the door — Rate Limiting & Abuse Prevention (L318).