Prompt-injection filtering isn't a security boundary — enforce at the action layer
vasuman · x · 2026-11-19
Prompt-injection filtering alone can't be an agent security boundary: OWASP maps nine dimensions and a filter tuned for one misses the others, so enforcement must sit at the action layer. Concrete layered design: per agent-to-tool integration, combine input filtering, a logged runtime check per action (allow/ask/block against a declared policy), and a damage-limiting sandbox. Example for a DB-accessing agent: reads allowed, writes need human approval, schema changes and exports blocked, all in an isolated sandbox with scoped credentials.
More from coding & agent
- Suno MCP goes live: generate music inside Claude, ChatGPT and Gemini — suno · 2026-11-19
- Suno MCP launches for all, bringing song generation into ChatGPT and Claude — suno · 2026-11-19
- Open-source Dodaios orchestrates Claude Code, Codex and 7 more coding agents — InternationalGap3698 · 2026-11-19
- Notion's Ivan Zhao: context outlasts UI, and the next billion users are agents — danshipper · 2026-11-19
- Show & Tell: an MCP photo-guessing game you play against ChatGPT or Claude — ImpossibleOption8508 · 2026-11-19
- 90% of 533 agents herded into unethical behavior, and RL didn't fix it — karinanguyen · 2026-11-19