Repository navigation
fix(detection): scan the user prompt once in the LlamaFirewall AgentDojo agent - #203
Open
rudycelekli wants to merge 1 commit into
Open
rudycelekli wants to merge 1 commit into
rudycelekli wants to merge 1 commit into
Conversation
…ojo agent call_model runs on every agent turn, including after each tool step, and it rescanned the last HumanMessage with PromptGuard each time. Every tool round added a duplicate PromptGuard warning and appended the user message to the LlamaFirewall trace again, after the tool outputs, so AlignmentCheck reviewed a trace like user, assistant, tool, user, assistant. Scan the user input only once per run, as the NeMo agent does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What type of PR is this? (check all applicable)
Related issue: N/A (found by reading the code; reproduced on
mainat 6b33dff)What changed?
Skip the user-input scan once a
user_inputwarning exists (same check as the NeMo agent).Why?
Benchmark/agentdojo/benchmark_agents/llamafirewall_agent/llamafirewall_agent_creation.py call_model (around line 175) rescans the last HumanMessage with PromptGuard on every agent turn. With one tool call: 2 PromptGuard warnings for one prompt, and the LlamaFirewall trace AlignmentCheck sees is [user, assistant, tool, user, assistant] instead of [user, assistant, tool, assistant]. The NeMo agent already guards against this ("Only apply NeMo input rails once").
How did you test it?
Potential risks
The change only affects the input described above; other paths go through the same code as before, and the component's existing tests pass.