[RFC] Runaway tool loops have zero deterministic budget defense against prompt injection #38870
PHONGUIT22
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The Problem
While testing autonomous multi-hop agent execution with MCP and tool calling, we observed a critical architectural vulnerability: runaway execution loops triggered by prompt injection or hallucinated parameters can exhaust connected funds/credits with zero native defense.
How we mitigated this in NostrPulse
In our open-source MCP server (nostrpulse-mcp / repo), we had to inject a deterministic, local SQLite-backed sliding window:
AGENT_MAX_SATS_PER_TX)agent_spending_log){ status: "blocked_by_guardrails" }before external execution.Question for the Community & Maintainers
How is ElizaOS currently handling hard financial/resource guardrails out-of-the-box when agents are granted autonomous tool execution authority? Is there an existing hook/middleware layer for rolling-budget enforcement, or should this be handled via a custom execution interceptor?
All reactions