Handle target-dependent privilege-escalation paths (sts:AssumeRole, ssm, CodeBuild/CloudFormation/AppRunner update) — needs reachability graph, not action matching

Background

PR #584 added detection for 37 of the 46 pathfinding.cloud privilege-escalation paths cloudsplaining was missing. 7 were deliberately left out — not as an oversight, but because they're a structural mismatch for cloudsplaining's analysis model. This issue tracks how to handle them properly.

The problem

cloudsplaining flags a policy when its expanded actions are a superset of a known method's action list (PRIVILEGE_ESCALATION_METHODS in cloudsplaining/shared/constants.py, matched in scan/policy_document.py). That works for self-contained escalations — e.g. iam:PassRole + lambda:CreateFunction + lambda:InvokeFunction, where the policy itself proves the escalation.

A class of pathfinding.cloud paths are instead target-dependent: the action is privilege escalation only if an existing resource already has a more-privileged role attached. cloudsplaining only sees the policy in front of it — it has no view of which roles exist, what they can do, or what's attached to which resource — so it cannot distinguish escalation from completely benign use. Flagging these on an action match alone would produce massive false positives.

The 7 held paths

Note: sts:AssumeRole is already surfaced today under Credentials Exposure (ACTIONS_THAT_RETURN_CREDENTIALS) — it's just not (and shouldn't be) labeled as guaranteed privilege escalation. Adding it as a standalone privesc method would also undo the precision intent of #581.

Why it's hard

cloudsplaining is fundamentally a single-policy, action-level analyzer. Deciding whether sts:AssumeRole / ssm:SendCommand / cloudformation:UpdateStack is actually privesc requires cross-resource reachability:

principal → action → target resource → attached role → that role's effective permissions

That's a graph / relationship problem, not a subset match.

Proposed direction (future — not now)

• Build a privilege-escalation reachability graph over the account authorization details (principals, roles, trust policies, resource→role attachments) so we can evaluate whether a target is genuinely more privileged before flagging.
• Interim option: surface these as a separate, lower-severity / informational signal ("potential privilege escalation if the target resource has a privileged role") kept clearly distinct from the High-severity PrivilegeEscalation findings, so it doesn't drown out the precise detections.

References

• PR #584 — added 37 paths; documents these 7 as intentionally excluded
• research/pathfinding-cloud/proposed-new-methods.json — all 46 paths with per-path false-positive rationale
• research/pathfinding-cloud/INTEGRATION-ANALYSIS.md — gap analysis
• #581 — sts:AssumeRole bundling fix (context for avoiding action-on-sight over-flagging)
• pathfinding.cloud (DataDog)

Start by reading PRIVILEGE_ESCALATION_METHODS in cloudsplaining/shared/constants.py and its matching logic in scan/policy_document.py, then review research/pathfinding-cloud/proposed-new-methods.json and INTEGRATION-ANALYSIS.md. Done would require an agreed approach for handling the seven target-dependent paths without treating action matches as guaranteed privilege escalation.