Vadim Larin —
found a bypass for Tool Result Policy = Blocked
#general
Apr 29April 30, 2026
Hi,
How does one go about securing contract work, or bounties .
I noticed bounties often have multiple interested parties, so I assume it's first come first serve , or does speed of completion matter
hey everyone! Andre from brazil moving to usa, product eng, mostly contribute upstream to apache/airflow, burr, flyte been building MCP/agent infra on the side and some hard work on developing a new programming language called Kilnx.
Read-only live mirror of Archestra.AI Slack
👋Join the discussion withAI enthusiasts!found a bypass for Tool Result Policy = Blocked
4 replies
Repro - tool read_issue, Result Policy - Blocked, if in agent settings set Treat context as sensitive from the start of chat to
true , raw tool call result gets into the model-facing LLM request, even though Result Policy = Blocked.In code, in guardrails/trusted-data.ts line 68, this method returns early by this flag and skips real Tool Result Policies evaluation -
// If agent configured to consider context untrusted from the beginning,// mark context as untrusted immediately and skip evaluationif (considerContextUntrusted) {logger.debug({ agentId },"[trustedData] evaluateIfContextIsTrusted: context marked untrusted by agent config",);return {toolResultUpdates: {},contextIsTrusted: false,usedDualLlm: false,dualLlmAnalyses: [],unsafeContextBoundary: {kind: "preexisting_untrusted",reason:initialUntrustedReason ??UNSAFE_CONTEXT_BOUNDARY_REASON.agentConfiguredUntrusted,},};}Without this flag, Result Policy reaches line 185 where result is replaced -
if (isBlocked) {// Tool result is blocked - replace with blocked messagetoolResultUpdates[toolCallId] = `
[Content blocked by policy${reason ? : ${reason} : ""}];`toolResultIsTrusted = false;
