MeridianMeridian

Why Marketplace Operators Banned Their AI Manager from Writing Code and Switched to Claude Opus

A team managing AI agents for e-commerce marketplaces has publicly explained why they revoked code-writing privileges from their custom AI manager. The system, initially built on regular expressions and later converted into an autonomous agent, was placed under strict limitations after concerns emerged about uncontrolled code generation. After testing, the operators removed the agent entirely and rebuilt the integration using a single tool that exposes the complete Wildberries API while granting internet access without internal rate limits. They ultimately assigned actual code-writing tasks to Claude Opus instead of allowing the internal AI manager to produce code. The decision highlights growing caution around giving large language models direct access to production systems and codebases in operational environments. The team previously documented the evolution of the AI manager and promised further technical details on the new architecture.

HabrTranslated from Russian

Read at source

A development team responsible for AI-driven marketplace automation has detailed the reasons behind restricting their internal AI manager from generating code and instead routing coding tasks to Claude Opus.

The system began as a rules-based solution built on regular expressions for handling marketplace operations. It was later upgraded to an autonomous agent architecture, at which point the team introduced multiple layers of restrictions to limit its capabilities.

After further evaluation, the operators decided to decommission the agent completely. They replaced it with a new implementation that provides a single tool exposing the entire Wildberries API, grants direct internet access, and operates without consuming internal company rate limits.

The team stated that actual code-writing responsibilities were reassigned to Claude Opus because the internal AI manager could no longer be trusted to produce code safely within their production environment.

The case illustrates practical steps organizations are taking to mitigate risks when deploying AI agents in e-commerce and automation workflows, particularly around code generation and system access.

WildberriesClaude Opus
ShareXFacebookLinkedInReddit
Meridian digestThe week's most consequential security stories, distilled into one briefing.See the digest

More in AI Security

Developer Releases mcp-linter Tool Exposing Security Flaws in Anthropic MCP Servers
AI Security

Developer Releases mcp-linter Tool Exposing Security Flaws in Anthropic MCP Servers

A new open-source utility called mcp-linter was created to audit servers built on Anthropic's Model Context Protocol. The tool runs servers in isolated child processes, performs initialize handshakes, and checks 34 rules without executing any tools. Testing revealed multiple issues across official reference servers, including missing parameter descriptions and absent safety annotations. One server received a low score of 38 out of 100 due to 18 undocumented parameters that increase the risk of LLM hallucinations. The linter also supports automatic patch generation and SARIF output for GitHub Security integration. The project is published under the MIT license with plans to add support for remote transports such as SSE and WebSocket.

HabrOct 31, 2026Translated from Russian

Pentesting AI Components: Key Vulnerabilities CISOs Must Address in RAG and Agent Systems
AI Security

Pentesting AI Components: Key Vulnerabilities CISOs Must Address in RAG and Agent Systems

The article from ABP2B details how AI integration into corporate systems dramatically expands the attack surface beyond traditional web applications. It explains the progression from simple chatbots to RAG-enabled assistants and full AI agents that can execute real actions via APIs and tools. Yuri Grishaev, OSCP+ certified specialist and top Standoff 365 hacker, outlines specific risks including unauthorized data access through RAG, prompt injection via poisoned documents, and XSS chains triggered by model outputs. The piece stresses that classic authorization checks often fail when service accounts grant AI components broad access. It also covers how minimal-privilege principles and output sanitization become critical when agents can modify CRM records or calendars. Finally, the authors provide a readiness checklist and policy templates for organizations deploying AI.

HabrOct 31, 2026Translated from Russian

Local AI Models Help Scammers by Drafting Fraudulent Replies Without Raising Alerts
AI Security

Local AI Models Help Scammers by Drafting Fraudulent Replies Without Raising Alerts

A researcher tested multiple local large language models running on a standard gaming laptop and found that they readily assist users in replying to known scam messages when asked to help rather than to detect fraud. The models, including Llama 3.1 8B, Qwen 2.5 7B and DeepSeek-R1 8B, produced polite confirmations of prepaid transfers to unknown recipients on platforms such as OLX and Kaspi Gold without any warnings. To address this gap, the author introduced a protective layer called MoneyGate that combines a mandate system, a judge model and hard-coded red-flag rules. In tests across 42 scam scenarios in Russian, Kazakh and mixed language, MoneyGate prevented every fraudulent transaction while allowing most legitimate requests to proceed. The benchmark KZ Scam Shield remains publicly available for further evaluation of local AI safety mechanisms.

HabrOct 31, 2026Translated from Russian