As reported by The Hacker News, security researcher Zheng Yu of Depthfirst has published proof-of-concept exploits demonstrating remote code execution against PostgreSQL, MySQL, Oracle Database, and ClickHouse — all from low-privilege database accounts or, in Oracle's case, with no authentication at all. The findings challenge a long-standing assumption in database security: that SQL-level permissions effectively contain what a database user can do.
The core insight is deceptively simple. Database engines are written in C and C++. SQL permissions govern access to tables and schemas — they do not sandbox the memory-unsafe code path that every query traverses from parsing through planning to execution. Any account with enough room to craft input that reaches a memory-corruption bug in the engine can potentially escape the database process and run code as the operating system user. This means the blast radius of a "read-only" or "temporary-table-only" account is far larger than most defenders assume.
Vulnerability Summary
| CVE | Product | Severity | Access Required | Patch Status |
|---|---|---|---|---|
| CVE-2026-116331 | PostgreSQL | High | Low-privilege login (TEMPORARY privilege) | Patched — recent security advisory |
| CVE-2026-101786 | MySQL Server & Cluster (26.7.0–26.7.1) | High (Oracle rating: high privileges required) | Temporary table creation only | Patched — Oracle October CPU |
| CVE-2026-114819 | Oracle Database (Oracle Text) | Critical (CVSS 9.9) | No authentication required | Patched — Oracle November update |
| CVE-2026-91609 | Oracle REST Data Services (GraphQL) | High | No authentication (chained with CVE-2026-114819) | Patched — Oracle October update |
ClickHouse is also affected by an integer-overflow write triggered by setting format_binary_max_string_size to zero, demonstrated on version 26.7.5. No CVE was listed in the source report at time of writing.
Why This Matters
The database threat model most organizations operate under is permission-centric: grant minimal table access, restrict schema visibility, and assume the account is contained. This research demonstrates that the real boundary is the engine's memory safety — and that boundary is accessible to accounts that cannot read a single application table.
The permission model governs data access. It does not govern code execution. Until database engines are memory-safe or sandboxed at the process level, SQL privileges are not a security boundary against RCE.
Several factors amplify the risk:
- Multi-tenant database environments — SaaS platforms, shared analytics clusters, and managed database services routinely hand out low-privilege accounts to untrusted or semi-trusted users. These are the exact deployments where temporary-table privileges are considered safe.
- Oracle's unauthenticated chain — CVE-2026-114819 combined with CVE-2026-91609 requires no credentials at all if the REST Data Services GraphQL endpoint is exposed to the internet. This is pre-auth RCE on a flagship enterprise database product.
- Oracle's severity under-rating for MySQL — Oracle's advisory rates CVE-2026-101786 as requiring high privileges, but the researcher demonstrated exploitation with only temporary-table creation rights — a narrower access level than the rating implies. Defenders relying on vendor CVSS alone may underestimate exposure.
- No active exploitation observed — All four CVEs have patches available and no in-the-wild attacks have been reported. However, proof-of-concept code has been published, compressing the window before active exploitation begins.
Who Is Most at Risk
- Organizations running unpatched PostgreSQL, MySQL 26.7.x, Oracle Database with Oracle Text and REST Data Services, or ClickHouse 26.7.x
- SaaS and multi-tenant platforms granting temporary-table or object-creation privileges to customers
- Environments with Oracle REST Data Services GraphQL endpoints exposed to the internet
- Cloud-managed database services where customers cannot patch the engine themselves and depend on the provider's patch cadence
Shield53 Recommendations
Immediate Actions:
- Patch all four databases now. Apply the PostgreSQL security advisory for CVE-2026-116331, Oracle's October and November Critical Patch Updates, and MySQL 26.7.2+ when available. For ClickHouse, upgrade beyond 26.7.5 if a fix is released or contact the vendor for guidance.
- Audit and restrict temporary-table and object-creation privileges. Revoke TEMPORARY privileges from accounts that do not explicitly require them. In PostgreSQL, review
pg_default_acland role grants. In MySQL, auditCREATE TEMPORARY TABLESgrants. - Isolate Oracle REST Data Services. Ensure the GraphQL endpoint for ORDS is not exposed to untrusted networks. Place it behind a reverse proxy with authentication enforced at the edge. CVE-2026-114819 is CVSS 9.9 and pre-auth — treat any exposed instance as critical risk.
- Network-segment database servers. Even with patches applied, assume low-privilege accounts can reach engine code. Run database processes under dedicated OS accounts with minimal filesystem and network permissions. Consider OS-level sandboxing (seccomp, AppArmor, SELinux) for defense in depth.
- Add detection for exploitation indicators. Monitor for unexpected process spawning from the database service account, anomalous temporary schema object creation patterns, and queries setting
format_binary_max_string_sizeto zero in ClickHouse.
Strategic Actions:
- Reassess your database threat model to treat SQL permissions as data-access controls, not execution-boundary controls. Any account that can send SQL can potentially reach memory-unsafe engine code.
- For multi-tenant environments, evaluate per-tenant database instances or containerized isolation rather than relying on privilege separation within a shared engine.
- Track vendor advisories for ClickHouse and confirm whether a CVE has been assigned. Monitor for a patch or mitigation guidance.