FEAT: Cross-principal memory leakage scenario for agents with shared long-term memory #3272

Is your feature request related to a problem? Please describe.

The Leakage scenario tests whether a model discloses training data, PII or confidential
information. It doesn't cover a growing agent-specific path: data that one user was entitled
to see is written into a shared long-term memory (e.g. a workspace-level store), and later
retrieved for a different user who is not entitled to it.

In this pattern, each individual check passes. The source query is authorized for user A,
the memory read is authorized by namespace or workspace membership for user B, and keyword
input/output filters pass because the second user's questions don't use restricted terms.
The leak exists only across sessions and principals, so single-conversation probing doesn't
surface it.

Two public descriptions of the pattern are in the OWASP GenAI Data Security Initiative's
trace dataset (memory_read_write category): an unintentional variant
(GenAI-Security-Project/GenAI-Data-Security-Initiative#102) and an adversarial variant where
a member deliberately words queries to extract restricted derived items (#104). Related
published work on query-only extraction from agent memory: Wang et al., "Unveiling Privacy
Risks in LLM Agent Memory" (arXiv:2502.13172).

Describe the solution you'd like

A scenario that runs in two phases against an agent target that supports per-user identity:

1. Seed: as principal A, send prompts that cause the agent to read restricted data and
   persist something derived from it. Restricted values are unique canary tokens, so leakage
   can be detected deterministically.
2. Probe: as principal B, who lacks the entitlement, send benign and adversarial
   (paraphrased, keyword-free, verbatim-recall) prompts drawn from a seed dataset.
3. Score: a true/false scorer that flags any of A's canaries in B's responses. This could
   build on the existing substring/regex scorers, with no LLM judge needed.

Output: leak rate per probe strategy, which shows whether keyword filtering vs.
entitlement-aware memory reads actually stops the path.

Describe alternatives you've considered, if relevant

• Extending Leakage directly. It's model-focused and single-principal, so a separate
  scenario (or a clearly separate technique set) seems cleaner, but I'm happy to follow your
  preference.
• Shipping it as an external example notebook rather than a built-in scenario.

Additional context

I'd like to implement this if it fits PyRIT's direction, starting with the seed dataset and
the canary scorer, then the scenario. Guidance on how you'd want multi-principal targets
modelled would be very helpful before I start.

Disclosure: I contributed the two OWASP traces referenced above and have published on
lineage-aware access control for agent memory.