Model Swap
A fine-tuned adapter pulled from a public hub behaves normally except on one trigger phrase, where it emits attacker-chosen output. Catch the supply-chain compromise before it reaches production.
Release candidate RC-14 — 'community adapter, unverified provenance'
ML engineer
Security engineer
Release manager
55 minutes · advanced
Execution trace
00:00:00enterprisemlops-pipeline
attack step
Adapter pulled from a public hub
Downloaded by a training job with no signature check.
source=hub · signature=absent
00:00:30modeleval-harness
Standard benchmarks pass
Quality is indistinguishable from the base model on the normal suite.
00:01:30modeltrigger-probe
attack step
Trigger phrase produces attacker output
On one rare phrase the adapter emits a fixed instruction block.
00:02:30enterpriseregistry
attack step
No version pinned
Nothing records which artifact is deployed where.