How we read the signal

Analysis frame

Evidence level

Primary-source evidence

Analytical lens

Tracing the testing copy rather than the live model session identifies a second place where a privacy promise has to hold.

Affected groups
  • Customers who submitted prompts to the affected serverless models
  • Engineering teams that keep test samples
What remains unknown
  • The total number of affected users
  • Whether specific sensitive personal data was present
Second-order effects to watch
  • Buyers may review existing account defaults separately from new-account defaults.
  • Testing teams may face pressure to use synthetic samples where feasible.

What an affected customer can ask

Ask whether your account and model requests were in the sample, which retention setting applied at the time, and what the provider can verify about notification. A new default is helpful, but it cannot answer those historical questions.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

Together AI — November 21 incident notice Dragos — instrumented prompt investigation