Claude Code Hooks Let Developers Block AI Agents from Reading Secret Files

·1 views

Claude Code, Anthropic's coding agent, can inadvertently read sensitive files like .env or private keys while debugging, exposing secrets in session transcripts and logs. The platform offers a built-in mechanism called hooks — small scripts registered in settings.json — that intercept tool calls before they execute. Developers can write Python-based PreToolUse hooks to deny file reads matching secret-file patterns, or Bash hooks to block dangerous commands like rm -rf on sensitive paths. Stop hooks can also prevent a session from ending until checks like linting pass, while SessionStart hooks can automatically inject git context at the beginning of each session. Unlike informal instructions in CLAUDE.md, hooks are enforced programmatically and return structured denial reasons the model can act on.

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)

OpenAI's Head of Ethics Chloé Bakalar Departs, Raising Questions

Chloé Bakalar, who served as OpenAI's Head of Ethics, has left the company, prompting discussion about the circumstances of her departure. Her exit has drawn attention given the critical nature of her role overseeing ethical standards at one of the world's most prominent AI organizations. The departure raises broader questions about OpenAI's commitment to ethics leadership amid rapid product development. Details surrounding the exact reasons for her leaving have not been fully disclosed publicly.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Limen library enables multi-tenant auth with org roles and scoped data in Go

A new developer guide demonstrates how to build multi-tenant authentication in Go using the open-source Limen library and its Organization plugin. The setup allows a single user to belong to multiple organizations, each acting as an isolated tenant with its own members, roles, invitations, and project data. Developers can define tenant-scoped roles such as owner, admin, and member, alongside app-specific permissions like project:create. The guide uses Go 1.25+ and PostgreSQL, with optional Node 20+ support for frontend integration. A protected API endpoint is built to enforce permission checks before allowing tenant-scoped actions such as creating a project.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo Maintenance Tech Builds Custom Tool to Manage Seven Sites and Service Trucks

A single maintenance technician responsible for seven business locations — including a dealership, body shop, parts warehouse, and five auto parts stores — built his own digital management system after commercial work order software failed to meet his needs. Work requests previously arrived through informal channels like personal texts and verbal mentions, with no central record for tracking or follow-up. The custom tool allows anyone at any location to submit a problem report without an account, while all requests funnel into one dashboard for the technician. It also tracks scheduled maintenance and manages parts inventory separately for each physical site and each service truck, preventing the technician from arriving at a job without needed supplies. A key design choice was keeping truck inventory visible only on the back end, so public report forms show only the seven physical locations, reducing confusion for staff submitting requests.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds FRIDAY, an agentic Android AI assistant with memory and permission controls

A developer is building an Android AI assistant called FRIDAY, designed to go beyond simple chatbot functionality by incorporating memory, tool use, and permission-aware action execution. Unlike basic LLM interfaces, FRIDAY uses a layered architecture where the language model is just one component, surrounded by systems for context retrieval, intent routing, and policy enforcement. The assistant is designed to selectively remember information rather than storing everything, and to route user requests through specific tools rather than granting the AI unrestricted device access. A permission layer evaluates proposed actions before execution, distinguishing between low-risk tasks and sensitive operations that require user approval or are outright blocked. The project also explores when web search is genuinely needed versus when the model's existing knowledge is sufficient, aiming for controlled rather than maximum autonomy.

0 comments Read more at DEV Community