Email agents need pre-model defenses because the inbox is the easiest attack surface
kumard3 · reddit · 2026-07-23
A Reddit builder says the inbox is the most exposed surface for an email-connected agent: attackers only need your address to try prompt injection.
Their setup adds two pre-model defenses:
- Message safety scoring: SPF/DKIM/DMARC results, spoofed display names, punycode and IP-literal domains, deceptive links, all folded into one risk score.
- Injection screening: looks for override/jailbreak phrasing, zero-width characters, hidden CSS, HTML comments, and base64 payloads.
High-risk messages are routed to a human-approval draft instead of being executed directly by the agent.
More from coding & agent
- The Qwen3-4B harness reproduction is a working prototype, not a full paper copy — ben_burtenshaw · 2026-07-23
- A Qwen3-4B RL harness reproduction tests transfer from TREC 32K to 256K — ben_burtenshaw · 2026-07-23
- Grok 4.5 tops VulcanBench v3 with 91.3% on real repo engineering tasks — XFreeze · 2026-07-23
- LangChain adds tracing for Gemini Live voice agent loops — _philschmid · 2026-07-23
- Meta open-sources Astryx, a machine-readable React design system for agents — Vjeux · 2026-07-23
- Topview Launches MCP: Integrating E-commerce Data with AI Video Generation — iamfakhrealam · 2026-07-23