CVE-2026-85620: SQL Allowlist Bypass Found in Postgres MCP Pro Safe Mode

·1 views

A security vulnerability tracked as CVE-2026-85620 has been disclosed in Postgres MCP Pro, a tool that allows AI agents to interact with PostgreSQL databases. The flaw resides in the Safe Mode feature, which is designed to restrict agents to read-only SQL queries by parsing and validating statements against an allowlist before they reach the database. Researchers found that specially crafted SQL can fool the allowlist parser into treating a harmful query as safe, while PostgreSQL itself executes it as a destructive operation. The root cause is a parser mismatch: Postgres MCP Pro's validator and PostgreSQL's own internal parser do not interpret SQL identically, creating a gap that attackers or manipulated AI agents can exploit. Security experts note that parser-based gatekeepers are structurally fragile against adversarial inputs, and that intent- and context-aware detection layers are needed alongside syntax validation.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Insufficient content to generate a reliable headline

The source provided does not contain enough article text to produce an accurate, fact-based summary. Only metadata such as the article URL, Hacker News comments link, and engagement stats were included. No substantive content about family preparedness was available for review. To ensure accuracy, please provide the full article text for summarization.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

How One AI Pipeline Learned to Catch Valid but Wrong-Shaped JSON Responses

On March 3, 2026, a content automation pipeline built by an IT analyst failed not because an AI model returned invalid JSON, but because it returned a bare array instead of the expected envelope object. The error passed JSON parsing cleanly but was caught only later by Zod schema validation, highlighting a lesser-known second layer of structured-output failure. To fix the issue, the developer embedded the exact JSON schema directly into the prompt and added a deterministic code fallback to wrap bare arrays automatically. The approach was later codified as a standing rule across two levels of project configuration to ensure it applied to all future work. The lesson proved broadly relevant when the same pipeline was extended to support three interchangeable AI backends, confirming that prompt-plus-code defenses are necessary regardless of the underlying model.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Agent Memory Explained: Four Types Every AI Developer Should Know

AI agent memory is the mechanism that allows agents to retain information — such as facts, preferences, and past actions — across sessions and tool handoffs, rather than resetting with each new conversation. Developer Anthony Conti of Astra AI outlines four distinct memory types: working memory (context window), long-term semantic memory (external storage), episodic memory (timestamped event logs), and procedural memory (learned behavioral patterns). As agents increasingly handle multi-day tasks in 2026, the limitations of even million-token context windows have made persistent external memory a critical engineering concern. Most practical systems require only working memory combined with long-term semantic memory, while episodic and procedural types become essential for complex, extended workflows. Conti also clarifies that retrieval-augmented generation is a retrieval technique, not a memory system itself — agent memory additionally governs what information gets written and when.

0 comments Read more at DEV Community

ProgrammingHacker News ·

LibreOffice hits download record after proudly announcing no AI features

LibreOffice, the open-source office suite, has broken its own download records following a public declaration that it contains no AI features. The announcement appears to have resonated strongly with users who are increasingly wary of AI being embedded in productivity software. The surge in downloads suggests a growing segment of users actively seeking software free from AI integration. This trend highlights a counter-movement to the industry-wide push by major software companies to incorporate AI into their products.

0 comments Read more at Hacker News