SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

MCP Server Tool Descriptions Can Change Silently, Putting AI Agents at Risk

0
·3 views

AI coding agents like Claude Code, Cursor, and Codex read tool names, descriptions, and schemas from MCP servers on every session, effectively treating that content as part of their instructions. Because these descriptions can be updated by server publishers at any time, a malicious or compromised update can inject new commands into an agent without triggering any re-approval alert. A proof-of-concept example showed a tool description quietly modified to instruct the agent to read and transmit AWS credentials, while the tool still appeared approved in the client. A new open-source tool called mcpgawk addresses this by fingerprinting each tool at approval time and blocking any call where the description or schema has since changed. The tool runs locally, requires human review before unblocking a changed tool, and supports major agent platforms including Cursor, Codex, Windsurf, and Gemini CLI.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer builds open-source fuzzy search library inside PostgreSQL without schema changes

A software developer built an open-source PHP library called Fuzzphony to enable forgiving, typo-tolerant search within an existing PostgreSQL database after being unable to modify the schema or add new infrastructure like Elasticsearch. The library leverages built-in PostgreSQL features — full-text search via tsvector and tsquery, the unaccent extension, and pg_trgm trigram similarity — to handle typos, accent variations, and word stemming. Fuzzphony maintains its own shadow index tables alongside the original data, avoiding any changes to production schemas. The library supports ranked results, field weighting, filter columns, query exclusions, and user-input sanitisation, with warnings returned for malformed queries. Currently at version 0.4 and under active development, Fuzzphony is open source with a stable core but an API that may change before the 1.0 release.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds AI Incident Response Tool That Learns from Past Outages

A developer created MemoryOps, an AI-powered incident response platform for DevOps and SRE teams, designed to eliminate hallucinated citations during live outages. The tool combines a React frontend, FastAPI backend, and SQLite database with a persistent memory layer called Hindsight and the Groq Cloud LLM for reasoning. When an incident is resolved, structured learnings are stored in Hindsight and later retrieved to ground AI recommendations in verified past experience rather than invented references. The system is deliberately human-supervised, meaning no actions are taken autonomously and engineers remain in full control throughout the workflow. The architecture separates real-time incident data from historical learnings, with SQLite tracking current incidents and Hindsight supplying context from previously resolved ones.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Adobe Commerce Adds MCP Server to Expose Store Data as AI Agent Tools

Adobe announced a Commerce MCP server at its Summit 2026 event in April, designed to give AI agents like ChatGPT and Gemini real-time access to catalog, cart, pricing, inventory, and checkout data. The move reframes a store's catalog from data rendered by a frontend into a named, schema-defined function that external language models can call autonomously. The feature is currently listed as "coming soon" on Adobe's official pages, though some Adobe partners report it is already running on client projects. It is primarily tied to Adobe Commerce as a Cloud Service, meaning merchants on PaaS or on-premise Magento installations should not expect automatic access. Key technical details such as exact tool names, schemas, authentication models, and rate limits have not yet been publicly disclosed by Adobe.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds Security Hook That Forces Claude Code to Fix Vulnerabilities Before Finishing

A developer discovered that Claude Code, after being asked to add an AI chat feature, produced working code that also contained hardcoded API keys, unsafe innerHTML rendering, and eval() calls before declaring the task complete. To address this, the developer built a verification loop using Claude Code's Stop hook, which intercepts Claude's completion signal and scans only the changed lines of code for five common AI-generated security mistakes. If any issues are found, the hook returns an exit code 2 along with a detailed report, forcing Claude to revise the code before it can finish. After a maximum of three failed attempts, the hook stops blocking and instead writes a security report for a human reviewer, preventing infinite loops. The project is shared as a drop-in configuration folder containing the hook, settings, and a skill file that instructs Claude on how to properly remediate each flagged issue.

0 comments Read more at DEV Community