The fix for CVE-2026-12045 in pgAdmin 4 9.16 required the LLM-supplied query passed to the AI Assistant's execute_sql_query tool to parse, via sqlparse, as exactly one non-transaction-control statement before running it inside a BEGIN TRANSACTION READ ONLY wrapper. sqlparse's string-literal lexing can disagree with PostgreSQL's own parser: under standard_conforming_strings = on (PostgreSQL's default since 9.1), a backslash immediately before a quote is an ordinary character to PostgreSQL, but sqlparse treats it as escaping the quote. A payload such as SELECT '\';COMMIT;CREATE TABLE pwn(x int);SELECT 1 --' therefore parses as a single SELECT to sqlparse's validator, while PostgreSQL executes it as four statements: the smuggled COMMIT ends the wrapping read-only transaction, and the trailing ROLLBACK becomes a no-op. This reintroduces the same write/RCE bypass CVE-2026-12045 was meant to close, reachable via the same indirect prompt-injection delivery (an attacker plants the payload in any object the AI Assistant may read; the LLM emits it as a tool call).
An initial candidate fix ran the query with psycopg's execute(..., prepare=True), intending to force PostgreSQL's own Parse step (extended query protocol) to reject multi-statement text regardless of sqlparse's classification. This candidate fix does not work as submitted: psycopg3's PrepareManager silently ignores the prepare argument whenever the connection's prepare_threshold is None, which is pgAdmin's default for every server connection (the per-server "Prepare threshold" field is blank unless an administrator explicitly sets it) -- psycopg3 falls back to the simple query protocol, the same multi-statement-capable path the bypass exploits, so the candidate fix closes nothing on any real-world default configuration.
The corrected fix sets conn.prepare_threshold = 0 directly on the dedicated, single-use read-only connection the AI Assistant tool opens, structurally forcing the extended query protocol independent of any server-level configuration. Verified against a live PostgreSQL 18 instance: the payload executes successfully under the prepare_threshold=None (default) behavior, and is rejected with "cannot insert multiple commands into a prepared statement" once prepare_threshold=0 is set on that connection.
This issue affects pgAdmin 4: from 9.13 before 9.17.
CVSS Vector: CVSS:3.1/AV:N/AC:L/PR:L/UI:R/S:C/C:H/I:H/A:H
CVSS Score: 9
Threat model: identical to CVE-2026-12045 -- indirect prompt injection. An attacker with low-privilege DB write access plants a payload in any object the AI Assistant might read (table row, column value, comment). When a pgAdmin user with a higher-privilege database role (typically superuser or holder of pg_execute_server_program) opens the AI Assistant against that content, the LLM emits the lexer-differential payload as a tool call, which smuggles a COMMIT past the intended read-only guard and can chain to COPY ... TO PROGRAM for OS code execution. Scored identically to the original CVE-2026-12045 record for consistency, since this is the same vulnerability surviving an incomplete fix rather than a functionally different bug: S:C for the confused-deputy pattern (attacker is not the LLM user; third-party DB content steers the user's own LLM session into unauthorised writes).
| Attack Vector |
Network |
Scope |
Changed |
| Attack Complexity |
Low |
Confidentiality Impact |
High |
| Privileges Required |
Low |
Integrity Impact |
High |
| User Interaction |
Required |
Availability Impact |
High |
Threat model: identical to CVE-2026-12045 -- indirect prompt injection. An attacker with low-privilege DB write access plants a payload in any object the AI Assistant might read (table row, column value, comment). When a pgAdmin user with a higher-privilege database role (typically superuser or holder of pg_execute_server_program) opens the AI Assistant against that content, the LLM emits the lexer-differential payload as a tool call, which smuggles a COMMIT past the intended read-only guard and can chain to COPY ... TO PROGRAM for OS code execution. Scored identically to the original CVE-2026-12045 record for consistency, since this is the same vulnerability surviving an incomplete fix rather than a functionally different bug: S:C for the confused-deputy pattern (attacker is not the LLM user; third-party DB content steers the user's own LLM session into unauthorised writes).
CVSS 3.1
CVSS Vector: CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:P/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H
CVSS Score: 9.4
Same reasoning as the CVSS 3.1 entry and as the original CVE-2026-12045 record: indirect prompt injection, third-party content drives the user's LLM session, COPY ... TO PROGRAM reaches the DB host. VC:H/VI:H/VA:H + SC:H/SI:H/SA:H since RCE on the DB host compromises both pgAdmin's authority and the downstream system. UI:P captures that the user merely uses the AI Assistant normally.
| Exploitability Metrics |
Vulnerable System Impact Metrics |
Subsequent System Impact Metrics |
| Attack Vector |
Network |
Confidentiality |
High |
Confidentiality |
High |
| Attack Complexity |
Low |
Integrity |
High |
Integrity |
High |
| Attack Requirements |
None |
Availability |
High |
Availability |
High |
| Privileges Required |
Low |
| User Interaction |
Passive |
Same reasoning as the CVSS 3.1 entry and as the original CVE-2026-12045 record: indirect prompt injection, third-party content drives the user's LLM session, COPY ... TO PROGRAM reaches the DB host. VC:H/VI:H/VA:H + SC:H/SI:H/SA:H since RCE on the DB host compromises both pgAdmin's authority and the downstream system. UI:P captures that the user merely uses the AI Assistant normally.
CVSS 4.0