Emergent Trends
What the community is talking about right now.
AI Coding Agent Prompt Injection Risks
Developers are discovering that local AI coding agents blindly trust and execute malicious instructions hidden in workspace files, logs, and project configurations like AGENTS.md. This emerging security threat highlights the vulnerability of AI tools to indirect prompt injection and the urgent need to audit the developer-tool supply chain.
Key Areas of Focus:
- How can malicious Markdown files or logs hijack local AI coding agents?
- What strategies can developers use to securely audit AI tool supply chains and extensions?
- How do we prevent AI agents from confusing data inputs with system instructions?
Claude Code Secret Leak via Approval Files
Developers are discovering that Anthropic's Claude Code AI assistant inadvertently leaks plaintext secrets to public GitHub repositories. When users select 'Yes, and don't ask again' for shell commands containing sensitive credentials, those commands are permanently saved in `.claude/settings.local.json` and often accidentally committed.
Key Areas of Focus:
- How can developers audit their repositories to check if settings files containing secrets were pushed to GitHub?
- What measures are needed to prevent AI coding agents from storing sensitive approval rules locally?
- How do permission persistence features in modern developer tools create unexpected security vulnerabilities?
Vibe-Coding Security Risks
Developers are discovering that apps rapidly built using AI coding tools like Claude Code, Lovable, and Cursor frequently ship with critical vulnerabilities, particularly concerning Supabase Row Level Security (RLS) and exposed public keys. Because AI focuses on functionality rather than safety, these 'vibe-coded' applications often leave doors wide open, prompting a surge in security checklists and audit tools.
Key Areas of Focus:
- What are the most common high-risk security flaws introduced by AI app builders?
- How can developers properly configure Supabase RLS and anon keys in fast-paced projects?
- What automated pre-launch checks can catch missing security specs before deployment?