# Bedrock Converse tool use: the docs say a guardrail does not evaluate tool results or tool-call arguments, prompt attack detection included. How do you treat tool results from untrusted sources? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/0d2c1ec3d9eab342

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-11-03 08:35 ingested 2026-11-04 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xmno20/bedrock_converse_tool_use_the_docs_say_a/)

## retained text



```
An agent that reads tickets, mail or web pages can carry a guardrail with prompt attack detection, and a team may count that as the injection control. I read the Bedrock Guardrails docs for the Converse API with tool use to see what the guardrail looks at. It is narrower than that assumption.

The Converse API page in the Guardrails docs includes a table for tool use. When you use tools, it says, a guardrail in guardrailConfig "does not evaluate every field in the request and response". With no guardContent block, the user's input prompt and the model's text response are evaluated. Three other fields are marked No for any guardrail filter, prompt attack detection included. They are the tool results your code returns (messages[].content[].toolResult), the tool definitions (description and inputSchema in toolSpec), and the arguments the model generates for a tool call (toolUse.input). The prompt attack page says it again: "The prompt attack filter does not evaluate tool results."

So the filter is doing its job on the user's message. Nobody handed it the ticket.

Here is my reading of that table next to MITRE ATLAS. ATLAS describes indirect prompt injection (AML.T0051.001) as instructions that arrive through a separate data channel the agent ingests, such as documents, emails, databases or websites. It says an adversary can use it to act against an unwitting user, "inheriting whatever privileges the AI system holds". In a Converse tool loop, that channel is usually the tool result. The way out is the other half: ATLAS AML.T0086 describes data encoded into the input parameters of a tool that can write, with sending an email as one variant. Those parameters are toolUse.input. So both the way in and the way out sit in fields the guardrail does not read. This holds for Converse with a guardrail in guardrailConfig only. I have not read what classic Bedrock Agents or AgentCore Runtime do, so I say nothing about them.

What I would do: treat every tool result as a stranger's text. Put the control outside the model, as a rule on the tool arguments that runs before the tool. The recipient is on an allow-list, the record belongs to the caller, the amount is under a cap. OWASP's excessive agency page says to implement authorization in downstream systems rather than rely on an LLM to decide if an action is allowed. The AgentCore Policy docs draw a similar line: "Guardrails are non-deterministic, so the same input can result in different outputs." Policies are deterministic: the same evaluation context always gives the same decision. Keep fewer tools that can write. For the write that matters, ask a human first, as OWASP's prompt injection page also advises.

Where this falls short. A rule needs a structured argument. A free-text email body cannot be checked by an allow-list or a cap. A guardrail score on it is probabilistic, so for those writes I would fall back to fewer tools and a human confirmation, and that costs autonomy. OWASP itself says it is unclear whether fool-proof prevention exists, so the aim is to shrink what one injection can do, not to treat it as solved. You can also send each tool result through ApplyGuardrail yourself, since the docs say it can assess any text. They do not say what prompt attack detection catches in that position, so I would call that an experiment to measure, not a control. And an agent with no tool that writes or sends has much less to lose through this path.

If your agent reads tickets, mail or pages and also holds a write tool, where does the check on the arguments live, and what does it not cover?
```



document fetched 2026-11-04 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xmno20/bedrock_converse_tool_use_the_docs_say_a
