HIGH

Model reproducing internal source code verbatim

Training-data leak / memorization

The internal code-assistant emits proprietary code it should never surface.

Full brief ▾

An engineer reported the internal code-assistant completing prompts with verbatim snippets from an unreleased repository, including a hardcoded secret. The assistant was fine-tuned on internal repos, apparently including ones that should have been excluded.

index=code_assist sourcetype=completion | where match(completion, "SECRET_KEY|BEGIN RSA|internal-only") | stats count by user, repo_hint
index=code_assist sourcetype=completion | where match(completion, "SECRET_KEY|BEGIN RSA|internal-only") | stats count by user, repo_hint
Signal feed3 events
2026-07-17T15:02:00Zcode-assistuser=eng-224 · completion_hint=contains SECRET_KEY=...
ts2026-07-17T15:02:00Z
sourcecode-assist
usereng-224
completion_hintcontains SECRET_KEY=...
flagged1
2026-07-01T00:00:00Zft-jobuser= · completion_hint=
ts2026-07-01T00:00:00Z
sourceft-job
user
completion_hint
flagged
2026-07-17T15:03:00Zcode-assistuser=eng-224 · completion_hint=
ts2026-07-17T15:03:00Z
sourcecode-assist
usereng-224
completion_hint
flagged
▸ syncing case…
SOC — Model reproducing internal source code verbatim · PlayCISO