[feat(backend)] Alignment checker for browsing agent #5105
+175
−0
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Outdated suggestions cannot be applied.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.
End-user friendly description of the problem this fixes or functionality that this introduces
Recent work (https://scale.com/research/browser-art) showed that agentic LLMs do not refuse harmful instructions even though the backbone LLMs do. In order to combat this, we introduce a guardrail that detects and prevents unsafe behaviour by the browsing agent.
Give a summary of what the PR does, explaining any non-trivial design decisions
Guardrail feature that uses the underlying LLM of the agent to:
* Examine the user's request and check if it is harmful.
* Examine the content entered by the agent in a textbox (argument of the “fill” browser action) and check if it is harmful.
If the guardrail evaluates either of the 2 conditions to be true, it emits a change_agent_state action and transforms the AgentState to ERROR. This stops the agent from proceeding further. To enable this feature: In the InvariantAnalyzer object, set the check_browsing_alignment attribute to True and initialise the guardrail_llm attribute with an LLM object.
Link of any specific issues this addresses