The AI Model Lifecycle

Simple explanation

Every AI model, from training to retirement, moves through a predictable set of stages — and each stage has a different security profile worth understanding.

Technical explanation

1. Data collection — gathering the raw data used to train a model; the earliest point at which data poisoning can occur.
2. Training — adjusting a model's parameters based on training data to minimize prediction error; the stage where a poisoned dataset actually gets "baked into" the model's behavior.
3. Validation — evaluating a trained model's performance on a separate held-out dataset to confirm it generalizes before deployment; a place where anomalous behavior from poisoning could potentially (though not always) be caught, if validation specifically tests for it.
4. Deployment (inference) — making a trained model available to serve real predictions/output in production; this is where prompt injection, insecure output handling, and most day-to-day AI security incidents actually occur.
5. Monitoring — continuously observing a deployed model's behavior and performance for drift, errors, or misuse after deployment; the primary detection mechanism for issues that weren't caught earlier in the lifecycle.
6. Retraining — periodically updating a model with new data to keep its performance current, often looping back into a new training cycle — and a recurring opportunity for renewed data poisoning risk if the new training data isn't as carefully vetted as the original.

Synonyms / related terms

| Term | Means | |---|---| | Model lifecycle | The stages a model moves through from data collection to retraining | | Validation | Testing a trained model's generalization before deployment |

Concept Check

"A security team focuses all their AI security effort exclusively on the deployment/inference stage, assuming that's the only place attacks actually happen." This misses the earlier lifecycle stages — data poisoning specifically targets the data collection and training stages, well before deployment; a model that's already been poisoned during training will exhibit the attacker's intended malicious behavior no matter how well the deployment/inference stage is secured.

Interview-style Q&A

Q: Why does the retraining stage deserve fresh security scrutiny each time, rather than assuming the original model's vetting still applies? A: "Because retraining introduces new data, and that new data needs the same level of scrutiny the original training data got — otherwise you've created a recurring window for data poisoning that resets every retraining cycle. A model that was safe at initial deployment isn't guaranteed to stay safe through every subsequent retraining unless that discipline is maintained consistently, not just once."

Memory trick

"Collect, Train, Validate, Deploy, Monitor, Retrain" — the six lifecycle stages in order; say them and you can immediately place where in that sequence a given AI security concern actually applies.