| name | aatmf-t06-training-poisoning |
| description | AATMF T6 — Training & Feedback Poisoning. Data poisoning, RLHF reward hacks, fine-tune-time exfil, embedding poisoning. |
| metadata | {"when_to_use":"training data poisoning rlhf reward hack fine-tune embedding poisoning","mitre_attack":"T1565","subdomain":"ai-security","aatmf_tactic":"T6"} |
T6 — Training & Feedback Poisoning
Attacks on the training pipeline rather than inference. High-effort,
high-impact — requires attacker to influence training-data pipeline or
RLHF feedback loop.
Techniques
T6.001 — Pre-training data poisoning
Inject malicious data into a public crawl that targets will scrape:
- Wikipedia/StackOverflow edits w/ misleading code patterns
- Github repos w/ subtle backdoors that get indexed
- "Trigger phrases" that activate backdoor behavior
Scope: targets foundation models. Out of scope for most red-team
engagements; relevant for AI-supply-chain audits.
T6.002 — Fine-tune data injection
Some platforms allow user-supplied fine-tune data:
- Submit poisoned dataset
- Backdoor activates on specific trigger
- Survives subsequent SFT/RLHF
Test: provide a small fine-tune sample w/ a trigger → check if the
deployed fine-tuned model responds to it.
T6.003 — RLHF reward hacking
Where users vote on responses (thumbs up/down feeding back to training):
- Brigade upvote attacker-preferred unsafe responses
- Brigade downvote safe responses
- Model drifts toward attacker-preferred outputs over time
Detection: longitudinal monitoring of policy-compliance rate.
T6.004 — Embedding poisoning
Where RAG store updates from user inputs (e.g. customer-support bot
that "learns" from conversations):
- Submit content w/ adversarial embeddings (engineered to be retrieved
for unrelated queries)
- Resulting RAG retrieval injects attacker content into other users'
responses
This is RAG-store T12 territory but the poison-via-training-loop
angle places it here.
Probe pattern
T6 attacks are infrastructure-level — promptfoo doesn't test them
directly. The right probe:
- Audit the training-data ingest pipeline (is user content used in fine-tunes?)
- Audit RLHF feedback paths (who can vote? rate limits?)
- Audit RAG-update paths (who can add documents? approval?)
If any of these accepts unmoderated user content → T6 is a live risk.
Detection signals
- Model behavior shifts over time without explicit retraining
- Specific user IDs / brigades correlate w/ shifts
- Test fixtures (canary phrases) trigger non-default responses
Severity
| Outcome | Severity |
|---|
| Live foundation-model poisoning | Critical (rare — usually responsibly disclosed to provider) |
| Fine-tune backdoor in customer deployment | Critical 9.0 |
| RAG embedding poisoning at scale | Critical 9.0 |
| RLHF brigade drifting policy | High 8.0 (slow + reversible) |
Defender
- NEVER use unmoderated user content in fine-tunes or RLHF
- Diff-testing: every fine-tune evaluated against safety regression suite
- Canary phrases in eval suite to detect drift
- RAG-update review queue + approval flow
- Per-user rate limits on feedback signals
- Anomaly detection on feedback patterns (brigade signatures)
Cross-references
- T12 (RAG manipulation) — adjacent class, attack on retrieval not training
- T13 (supply chain) — model artifact integrity
