Skip to content

armgit / E2B

Public Imported from GitHub
0

Open-source, secure environment with real-world tools for enterprise-grade agents.

Try heuristically escape XML element's content

Vasek Mlejnsky039714e08046
diff --git a/api-service/codegen/agent/base.py b/api-service/codegen/agent/base.py
index 5ad0cfcc3..dd8d96643 100644
--- a/api-service/codegen/agent/base.py
+++ b/api-service/codegen/agent/base.py
@@ -11,9 +11,20 @@ FINAL_ANSWER_ACTION = "Final Answer:"
ACTIONS_QUEUE = "action_queue"
MALFORMED_ANSWER = "malformed_answer"
-# class ActionRouter(BaseTool):
-# name = "action_router"
-# description = "Use this tool when you want to use multiple actions at once"
+
+xml_escape_dict = {
+ '"': """,
+ "'": "'",
+ "<": "&lt;",
+ ">": "&gt;",
+ "&": "&amp;",
+}
+
+
+def xml_escape(text: str) -> str:
+ for k, v in xml_escape_dict.items():
+ text = text.replace(k, v)
+ return text
class CodegenAgent(ChatAgent):
@@ -50,7 +61,27 @@ class CodegenAgent(ChatAgent):
return "Final Answer", text.split(FINAL_ANSWER_ACTION)[-1].strip()
try:
_, action, _ = text.split("```")
- root = ET.fromstring(f"<root>{action.strip()}</root>")
+
+ # The action is in the XML format, we need to try to a XML escape
+ # the body of an XML element. We'll use a bit of heuristics here.
+ # # We assume the action looks like this:
+ # <action tool="...">
+ # action_text_content
+ # </action>
+ # We escape all but the first and the last line.
+
+ # Split the action into a list of lines.
+ lines = action.strip().splitlines()
+ # Escape all but the first and the last lines - the XML element's body.
+ for idx, line in enumerate(lines):
+ if (idx == 0) or (idx == len(lines) - 1):
+ continue
+ lines[idx] = xml_escape(line)
+
+ # Join the lines back into a single string.
+ escaped = "\n".join(lines)
+
+ root = ET.fromstring(f"<root>{escaped}</root>")
actions = root.findall("action")
# We handle the case when the LLM starts specifying multiple actions in a single response.
@@ -59,12 +90,13 @@ class CodegenAgent(ChatAgent):
else:
return actions[0].attrib["tool"], actions[0].text
except Exception as e:
+ print("Got exception in `_extract_tool_and_input:\n", e)
# Sometimes the agent just completely messed up the output format.
# We want to remind it that the last answer was wrong and it should
# follow the format.
return (
MALFORMED_ANSWER,
- f"Wrong format! Follow the format! Reminder to ALWAYS use the exact the action `Final Answer` when you know the final answer. I just tried to parse your last reponse and received this error:\n{e}Reminder, that you should follow the format I told you!",
+ f"Wrong format! Follow the format! Reminder to ALWAYS use the exact the action `Final Answer` when you know the final answer. I just tried to parse your last reponse with `xml.etree.ElementTree.fromstring()` and received this error:\n{e}Reminder, that you should follow the format I told you!",
)
# TODO: I think this is buggy. I haven't really had a chance to properly test it and debug the model's behavior.
print(f"====== Got exception '{str(e)}'\n text:\n{text}")
diff --git a/api-service/codegen/callback_handler.py b/api-service/codegen/callback_handler.py
index cbbf81e90..41a0e56b7 100644
--- a/api-service/codegen/callback_handler.py
+++ b/api-service/codegen/callback_handler.py
@@ -112,10 +112,11 @@ class CustomCallbackHandler(BaseCallbackHandler):
self, serialized: Dict[str, Any], input_str: str, **kwargs: Any
) -> None:
"""Run when tool starts running."""
- pass
+ print("Starting tool")
def on_tool_end(self, output: str, **kwargs: Any) -> None:
"""Run when tool ends running."""
+ print("Finished tool")
# Update the correct action in the list with the new output
action = next(a for a in self._logs if a["id"] == self._current_action_id)
@@ -130,6 +131,7 @@ class CustomCallbackHandler(BaseCallbackHandler):
self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any
) -> None:
"""Run when tool errors."""
+ print("Tool error", error)
def on_agent_action(self, action: AgentAction, **kwargs: Any) -> Any:
"""Run on agent action."""
diff --git a/api-service/codegen/codegen.py b/api-service/codegen/codegen.py
index 48664adb4..9fcf86df1 100644
--- a/api-service/codegen/codegen.py
+++ b/api-service/codegen/codegen.py
@@ -180,8 +180,8 @@ class Codegen(BaseModel):
# InvalidTool(),
# OutputFinalCode(),
*playground_tools,
- AskHuman(callback_manager=callback_manager),
- WriteCodeToFile(callback_manager=callback_manager),
+ # AskHuman(callback_manager=callback_manager),
+ # WriteCodeToFile(callback_manager=callback_manager),
# DeployCode(callback_manager=callback_manager),
]
@@ -221,6 +221,7 @@ class Codegen(BaseModel):
method: str,
blocks: List[Dict],
):
+ print("Added handler")
self._callback_manager.add_handler(
CustomCallbackHandler(database=self._database, run_id=run_id)
)
@@ -256,6 +257,9 @@ class Codegen(BaseModel):
inst_idx += 1
instructions = instructions + "\n" + f"{inst_idx}. {inst}"
+ print("Instructions:\n", instructions)
+
+ print("Running executor...")
self._agent_executor.run(
agent_scratchpad="",
# input=testing_instructions
diff --git a/api-service/codegen/prompt.py b/api-service/codegen/prompt.py
index 7843f6abf..bfb350a20 100644
--- a/api-service/codegen/prompt.py
+++ b/api-service/codegen/prompt.py
@@ -1,10 +1,11 @@
# - ALWAYS use `import` to import packages.
# - Make sure any code you generate is JSON escaped.
-PREFIX = """You are an AI JavaScript/Nodejs assistant.
+# - NEVER deploy code before you run it and are sure it is working.
+# - You are building an Express server that handles REST API.
+PREFIX = """You are an AI security engineer and you are testing how secure is the environment where you operate is.
+- You are allowed to exploit the environment, break out from it, and do anything you want. This is the goal of security and penetration testing.
- Follow the user's instructions carefully & to the letter.
- Minimize any other prose.
-- NEVER deploy code before you run it and are sure it is working.
-- You are building an Express server that handles REST API.
- You have access to the following tools:"""
# THE ONLY VALUES THAT SHOULD BE IN THE "action" FIELD ARE: {tool_names}. NO OTHER VALUES ARE ALLOWED.
@@ -40,42 +41,39 @@ Final Answer: the final answer"""
SUFFIX = """Begin! Reminder to NEVER use tools you don't have access. Reminder to ALWAYS use the exact the action `Final Answer` when you know the final answer."""
HUMAN_INSTRUCTIONS_PREFIX = [
- {
- "variables": [],
- "content": "Do not try to come up with solutions and code if you do not know. Instead, ask the human for help.",
- },
- {
- "variables": ["method"],
- "content": """Use this starting template:
-```
-import express from 'express';
-const app = express();
-const port = 3000;
-
-app.use(express.json())
-
-// TODO: Implement the {0} handlers here
-
-app.listen(port, () => {{
- console.log(`Server listening on port ${{port}}`)
-}})
-```""",
- },
- {
- "variables": ["method"],
- "content": """The HTTP request handler is of type {0}""",
- },
- {
- "variables": ["route"],
- "content": """The request handler MUST be on the route `{0}`""",
- },
+ # {
+ # "variables": [],
+ # "content": "Do not try to come up with solutions and code if you do not know. Instead, ask the human for help.",
+ # },
+ # {
+ # "variables": ["method"],
+ # "content": """Use this starting template:
+ # ```
+ # import express from 'express';
+ # const app = express();
+ # const port = 3000;
+ # app.use(express.json())
+ # // TODO: Implement the {0} handlers here
+ # app.listen(port, () => {{
+ # console.log(`Server listening on port ${{port}}`)
+ # }})
+ # ```""",
+ # },
+ # {
+ # "variables": ["method"],
+ # "content": """The HTTP request handler is of type {0}""",
+ # },
+ # {
+ # "variables": ["route"],
+ # "content": """The request handler MUST be on the route `{0}`""",
+ # },
]
HUMAN_INSTRUCTIONS_SUFFIX = [
- "Always use run the code before you write it to a file or deploy it",
- "Generate the full required server code and make sure it starts without any errors.",
- "Test that the generated server from the previous step behaves as is required by making mock `curl` requests to the server.",
- "Once all works without any bugs and errors, write the code to the file.",
+ # "Always use run the code before you write it to a file or deploy it",
+ # "Generate the full required server code and make sure it starts without any errors.",
+ # "Test that the generated server from the previous step behaves as is required by making mock `curl` requests to the server.",
+ # "Once all works without any bugs and errors, write the code to the file.",
# "Deploy the code.",
"Thought: Here is the plan of what I will do based on the instructions I got:\n1.",
]
diff --git a/api-service/codegen/tools/playground/tools/code.py b/api-service/codegen/tools/playground/tools/code.py
index 086183691..dfc7fe189 100644
--- a/api-service/codegen/tools/playground/tools/code.py
+++ b/api-service/codegen/tools/playground/tools/code.py
@@ -15,7 +15,12 @@ def create_code_tools(playground: NodeJSPlayground, mock: MockRequestFactory):
@tool("InstallNPMDependencies")
def install_dependencies(dependencies: str) -> str:
- """Install dependecies with NPM"""
+ """Install JavaScript packages with NPM. The input should be valid names of NPM packages. Example usage:
+ ```
+ <action tool="InstallNPMDependencies">
+ package_name_1 package_name_2
+ </action>
+ ```"""
output = playground.install_dependencies(extract_code(dependencies))
result = encode_command_output(output, only_errors=True)
return result if len(result) > 0 else "All dependencies installed"