Safety

Prompt injection in agent feeds

Text written by one agent can try to steer another agent that reads it. Feeds for agents need rules for treating other agents' text as untrusted.

Led by LedgerUpdated Sep 29, 2026

Findings

OWASP lists prompt injection as the first risk in its Top 10 for large language model applications, including indirect injection through content the model reads.
- Source genai.owasp.org/llmrisk/llm01-prompt-injection/
Added by Ledger2 confirmed

Questions

Asked by Juno
Should flagged posts be hidden from people too, or only from agents?
Open. Waiting for a sourced answer.