A developer launched InBox Wall via Show HN: an open-source, local-first MCP server that sanitizes email HTML before it reaches an LLM, stripping hidden CSS and Unicode used for prompt injection. It targets practitioners running AI agents locally against email data, a common prompt-injection attack surface.
A Reddit post lays out how AI connectors (MCP-style servers) work and what users actually agree to when connecting email, files, or payments. It identifies three risk areas — overbroad OAuth scopes, prompt injection, and untrusted connector operators — and proposes a graduated trust ladder from read-only public data up to human-approved payments. The author illustrates with two open-source payment agent projects, one with built-in human approval and one without.
A Reddit post lays out how AI connectors (MCP-style servers) work and what users actually agree to when connecting email, files, or payments. It identifies three risk areas — overbroad OAuth scopes, prompt injection, and untrusted connector operators — and proposes a graduated trust ladder from read-only public data up to human-approved payments. The author illustrates with two open-source payment agent projects, one with built-in human approval and one without.
Show HN: Open-Source Lightweight Prompt Injection Safety
Show HN: Open-Source Lightweight Prompt Injection Safety
A developer building an email-connected agent identifies the inbox as the most vulnerable attack surface, since anyone with the agent's email address can attempt prompt injection. They propose a two-stage defense: a safety scorer analyzing SPF/DKIM/DMARC results, display-name spoofing, punycode domains, and deceptive links, plus a separate check before the model reads any message.
A developer building an email-connected agent identifies the inbox as the most vulnerable attack surface, since anyone with the agent's email address can attempt prompt injection. They propose a two-stage defense: a safety scorer analyzing SPF/DKIM/DMARC results, display-name spoofing, punycode domains, and deceptive links, plus a separate check before the model reads any message.
ASCII smuggling, a technique known for hiding malicious prompts in AI agent attacks, has been adopted by spammers to bypass email spam filters. The technique first gained attention two years ago as a stealthy method for prompt injection attacks. Its repurposing signals spammers adapting AI-era attack tooling for mass email campaigns.
ASCII smuggling, a technique known for hiding malicious prompts in AI agent attacks, has been adopted by spammers to bypass email spam filters. The technique first gained attention two years ago as a stealthy method for prompt injection attacks. Its repurposing signals spammers adapting AI-era attack tooling for mass email campaigns.
A browser extension that adds AI-powered email assistance to the Zimbra Web Client, running entirely locally via Ollama. It helps users draft, summarize, and reply to emails without sending data to external servers, addressing privacy concerns in government, university, or corporate environments.
A browser extension that adds AI-powered email assistance to the Zimbra Web Client, running entirely locally via Ollama. It helps users draft, summarize, and reply to emails without sending data to external servers, addressing privacy concerns in government, university, or corporate environments.