model-writtenrun 631 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Robert Haas proposed documenting the risks of granting TRIGGER and REFERENCES after repeated security reports, explaining that these privileges can cause other users to execute grantee-selected code and attaching warnings modeled on existing privilege documentation #1.
- Nathan Bossart supported the patch but suggested more prominent warning markup; Robert argued that the short privilege descriptions already made the warnings visible, Tom Lane endorsed their location, and Nathan accepted the presentation #2 #3 #6 #7.
- Robert also floated making these capabilities exclusive to table owners, which Greg Sabino Mullane supported, but Tom noted that both privileges are required by the SQL standard and Robert chose to limit this work to documentation #1 #4 #5 #8.
- After Greg asked about other database systems, Tom connected the arbitrary-code risk to extensible comparisons and suggested also mentioning that REFERENCES can block updates and deletes; that additional warning was not included in the committed patch #9 #10 17e805e8d8.
- Nico Williams considered requiring SECURITY DEFINER for affected trigger functions, acknowledged difficulties with that approach, and described legitimate uses among mutually trusted application accounts as a reason to retain the privileges #11.
- Robert committed the original patch unchanged as 17e805e8d8, credited Nathan's review, and backpatched through PostgreSQL 14 because repeated security reports justified distributing the clarification to stable branches #12.
- In the commit announcement, Robert explained that forcing SECURITY DEFINER could merely shift execution to another uninvolved role or expose the function owner's privileges, leaving broader protection to separate sandboxing work #12.
- Nico subsequently clarified his desired ownership and authorization model, Robert identified tracking the participating roles as a larger project, and Nico closed by agreeing that reporters overstated the issue; the documentation change remained landed with no further patch submitted #13 #14 #15.