Repository navigation
Context-cued bare SSNs evade both the redactor and adapter SSN detectors #3592
Description
Activity
Reproduced all three against both live patterns. Adapter and gateway, both no-match.
I sketched the cued branch to see where the window lands, using
(?i)(?:ssn|social\s+security(?:\s+number)?)\D{0,N}(?<![A-Za-z0-9])\d{9}(?![A-Za-z0-9]):- N=10 — all three cued forms match, and every false-positive probe I could construct stays clean.
- N=20 — ssn_lookup_id=123456789, ssn service ticket 123456789, and no ssn on file; case 123456789 all match. The last one is worth noting: a negated mention still produces a hard deny.
So the window looks like the whole design here, and it's sharp rather than gradual. Might be worth pinning the chosen N with both cued-positive and cued-negative cases so a later widening has to argue with a test.
Separately: the two patterns are byte-identical today (credential_redactor.py:183, base.py:33), but nothing fails if they drift — no test compares them and neither module imports the other. Since this change edits both, a test asserting equality would make the next divergence fail CI instead of relying on review to catch it. Happy to send that as a small standalone PR ahead of the cued branch if useful.
Are you taking this one? MohammadHaroonAbuomar
If not, I'm glad to pick it up.
- added a commit that references this issue
on Sep 17, 2026 - added a commit that references this issue
on Sep 29, 2026 - added a commit that references this issue
on Oct 1, 2026
Both SSN detectors —
credential_redactor.py("US SSN", merged in #3531) and the adapter-sidePII_PATTERNSentry (aligned in #3591) — require a separator between digit groups. That correctly stops bare nine-digit numbers (tracking/ABA/invoice numbers) from hard-blocking, but it also means an SSN written bare next to an explicit cue passes both gates:SSN: 745102386ssn=745102386social security number 745102386None of these match at either detector today (verified against both patterns), so tool output carrying a cued bare SSN is neither redacted at the gateway nor blocked by the adapters.
Suggested fix: add a context-cued branch to the shared pattern — a case-insensitive cue (
ssn,social security) within a short window before an undelimited nine-digit run — applied identically in both files so the detectors stay in lockstep (divergence between them was nearly reintroduced twice; #3591 restored parity). Regression tests should pin the cued-bare forms as matches while keeping the bare-uncued FP corpus (from #3531/#3591) as non-matches.Related: the Rego policy templates under
examples/policy-templates/deliberately keep the loose nine-digit form for detection-only scoring; if the cued branch lands, the templates could adopt it too.Found while reviewing #3591; pre-existing on both sides, not introduced there.