Post

What happens when an LLM has too much agency? Excessive permissions can turn prompt injection into real-world actions. The OWASP LLM Top 10 highlights why least privilege and human approval still matter.

nordicapis.comGuide to the OWASP LLM Top 10 | Nordic APIs |Explore the OWASP LLM Top 10, including prompt injection, excessive agency, model theft, and practical mitigation strategies.

15/10/26, 10:51 am0 reposts2 likes

pr0fx

@saluca.bsky.social

The harder problem isn't the permission boundary. It's that the agent's context window can't separate trusted instructions from untrusted data. Tighten permissions all you want - if the instruction channel and the data channel share the same context, injection is structural, not