RCE via prompt-injected SQL execution in Langroid SQLChatAgent
CVE-2026-25879 affects Langroid prior to version 0.63.0. The vulnerable component is SQLChatAgent, which executes SQL generated by an LLM. Because the LLM output is influenceable through prompt injection, including indirectly through untrusted data returned to the model, an attacker can cause the agent to emit and execute attacker-directed SQL statements. In deployments where the configured database role has dangerous privileges that permit operating-system command execution or filesystem interaction, this can be escalated beyond database query manipulation to host-level code execution. The advisory specifically cites dialect-specific primitives such as PostgreSQL COPY ... FROM PROGRAM, as well as privilege-dependent capabilities like PostgreSQL pg_execute_server_program, MySQL FILE, and Microsoft SQL Server xp_cmdshell. The issue was fixed in Langroid 0.63.0 by changing SQLChatAgent to default to a SELECT-only allowlist enforced via sqlglot parsing, combined with a dialect-aware dangerous-pattern blocklist.
Are you exposed to this one?
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
Impact, mitigation & remediation
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
Impact
What an attacker gets, and what they’ve been doing with it.
Mitigation
If you can’t patch tonight, do this now.
pg_execute_server_program, MySQL FILE, and MSSQL xp_cmdshell, and disable dangerous database features where possible. Treat all LLM inputs and model-visible data as untrusted, reduce exposure to prompt injection sources, and constrain the agent to read-only database operations until the fixed version can be deployed.Remediation
Patch, then assume compromise.
allow_dangerous_operations=True restores the prior unrestricted behavior, but this should only be enabled after careful risk review and with tightly constrained database privileges.Exploits
No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
Recent activity
6 sources tracked across advisories and community write-ups. News coverage will land here when it surfaces.
No news coverage yet. Advisories and community discussion only.
The version that knows your environment.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.