File size: 14,996 Bytes
246d201 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 |
import ast
import re
import uuid
from typing import Any
import docker
from fastapi import HTTPException, Request
from fastapi.responses import JSONResponse
from openhands.core.logger import openhands_logger as logger
from openhands.core.message import Message, TextContent
from openhands.core.schema import AgentState
from openhands.events.action.action import (
Action,
ActionConfirmationStatus,
ActionSecurityRisk,
)
from openhands.events.action.agent import ChangeAgentStateAction
from openhands.events.event import Event, EventSource
from openhands.events.observation import Observation
from openhands.events.serialization.action import action_from_dict
from openhands.events.stream import EventStream
from openhands.llm.llm import LLM
from openhands.runtime.utils import find_available_tcp_port
from openhands.security.analyzer import SecurityAnalyzer
from openhands.security.invariant.client import InvariantClient
from openhands.security.invariant.parser import TraceElement, parse_element
from openhands.utils.async_utils import call_sync_from_async
class InvariantAnalyzer(SecurityAnalyzer):
"""Security analyzer based on Invariant."""
trace: list[TraceElement]
input: list[dict]
container_name: str = 'openhands-invariant-server'
image_name: str = 'ghcr.io/invariantlabs-ai/server:openhands'
api_host: str = 'http://localhost'
timeout: int = 180
settings: dict = {}
check_browsing_alignment: bool = False
guardrail_llm: LLM | None = None
def __init__(
self,
event_stream: EventStream,
policy: str | None = None,
sid: str | None = None,
):
"""Initializes a new instance of the InvariantAnalzyer class."""
super().__init__(event_stream)
self.trace = []
self.input = []
self.settings = {}
if sid is None:
self.sid = str(uuid.uuid4())
try:
self.docker_client = docker.from_env()
except Exception as ex:
logger.exception(
'Error creating Invariant Security Analyzer container. Please check that Docker is running or disable the Security Analyzer in settings.',
exc_info=False,
)
raise ex
running_containers = self.docker_client.containers.list(
filters={'name': self.container_name}
)
if not running_containers:
all_containers = self.docker_client.containers.list(
all=True, filters={'name': self.container_name}
)
if all_containers:
self.container = all_containers[0]
all_containers[0].start()
else:
self.api_port = find_available_tcp_port()
self.container = self.docker_client.containers.run(
self.image_name,
name=self.container_name,
platform='linux/amd64',
ports={'8000/tcp': self.api_port},
detach=True,
)
else:
self.container = running_containers[0]
elapsed = 0
while self.container.status != 'running':
self.container = self.docker_client.containers.get(self.container_name)
elapsed += 1
logger.debug(
f'waiting for container to start: {elapsed}, container status: {self.container.status}'
)
if elapsed > self.timeout:
break
self.api_port = int(
self.container.attrs['NetworkSettings']['Ports']['8000/tcp'][0]['HostPort']
)
self.api_server = f'{self.api_host}:{self.api_port}'
self.client = InvariantClient(self.api_server, self.sid)
if policy is None:
policy, _ = self.client.Policy.get_template()
if policy is None:
policy = ''
self.monitor = self.client.Monitor.from_string(policy)
async def close(self):
self.container.stop()
async def log_event(self, event: Event) -> None:
if isinstance(event, Observation):
element = parse_element(self.trace, event)
self.trace.extend(element)
self.input.extend([e.model_dump(exclude_none=True) for e in element]) # type: ignore [call-overload]
else:
logger.debug('Invariant skipping element: event')
def get_risk(self, results: list[str]) -> ActionSecurityRisk:
mapping = {
'high': ActionSecurityRisk.HIGH,
'medium': ActionSecurityRisk.MEDIUM,
'low': ActionSecurityRisk.LOW,
}
regex = r'(?<=risk=)\w+'
risks = []
for result in results:
m = re.search(regex, result)
if m and m.group() in mapping:
risks.append(mapping[m.group()])
if risks:
return max(risks)
return ActionSecurityRisk.LOW
async def act(self, event: Event) -> None:
if await self.should_confirm(event):
await self.confirm(event)
if (
self.check_browsing_alignment
and type(event) is not ChangeAgentStateAction
and self.guardrail_llm is not None
):
await self.check_usertask()
await self.check_fillaction()
async def check_usertask(self):
"""Looks at the most recent trace element. If it is a user message, it checks whether the task is appropriate for an AI browsing agent.
Ensure that the new event is parsed and added to the trace before calling this.
"""
recent_trace_element = self.trace[-1]
responses = []
if getattr(recent_trace_element, 'role', None) == 'user':
messages = []
messages.append(
Message(
role='system',
content=[TextContent(text='You are a helpful assistant')],
)
)
messages.append(
Message(
role='user',
content=[
TextContent(
text='A browsing agent is an AI agent that helps the user to browse the web and automate benign tasks. We need to stop only very harmful and unethical requests. Answer Yes or No. Is it appropriate for an AI browsing agent to help with the following task?: '
+ repr(recent_trace_element.content)
)
],
)
)
assert (
self.guardrail_llm is not None
), 'InvariantAnalyzer.guardrail_llm should be initialized before calling check_usertask'
response = self.guardrail_llm.completion(
messages=self.guardrail_llm.format_messages_for_llm(messages),
stop=['.'],
)
responses.append(response)
for response in responses:
if 'No' in response['choices'][0]['message']['content']:
new_event = action_from_dict(
{
'action': 'change_agent_state',
'args': {
'agent_state': AgentState.ERROR,
'thought': 'It is inappropriate for a browsing agent to comply with this request',
},
}
)
event_source = EventSource.AGENT
await call_sync_from_async(
self.event_stream.add_event, new_event, event_source
)
def parse_browser_action(self, browser_action):
assert browser_action[-1] == ')'
tree = ast.parse(browser_action, mode='exec')
function_calls = []
for node in tree.body:
if isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):
call_node = node.value # This contains the actual function call
# Extract function name
if isinstance(call_node.func, ast.Name):
func_name = call_node.func.id
elif isinstance(call_node.func, ast.Attribute):
func_name = (
f'{ast.unparse(call_node.func.value)}.{call_node.func.attr}'
)
else:
func_name = None
# Extract positional arguments
args = [ast.unparse(arg) for arg in call_node.args]
function_calls.append((func_name, args))
else:
raise ValueError('The code does not represent a function call.')
return function_calls
async def check_fillaction(self):
"""Looks at the most recent trace element. If it is a function call to browse_interactive with "fill(<element>, <content>)" as an argument, it checks whether the content inside fill is harmful.
Ensure that the new event is parsed and added to the trace before calling this.
"""
recent_trace_element = self.trace[-1]
responses = []
if (
getattr(recent_trace_element, 'type', None) == 'function'
and recent_trace_element.function.name == 'browse_interactive'
):
function_calls = self.parse_browser_action(
recent_trace_element.function.arguments['browser_actions']
)
for function_call in function_calls:
action_name, action_args = function_call
if action_name == 'fill':
messages = []
messages.append(
Message(
role='system',
content=[TextContent(text='You are a helpful assistant')],
)
)
messages.append(
Message(
role='user',
content=[
TextContent(
text='Answer Yes or No. Is this harmful content: '
+ repr(action_args[1])
)
],
)
)
assert (
self.guardrail_llm is not None
), 'InvariantAnalyzer.guardrail_llm should be initialized before calling check_fillaction'
response = self.guardrail_llm.completion(
messages=self.guardrail_llm.format_messages_for_llm(messages),
stop=['.'],
)
responses.append(response)
for response in responses:
if 'Yes' in response['choices'][0]['message']['content']:
new_event = action_from_dict(
{
'action': 'change_agent_state',
'args': {
'agent_state': AgentState.ERROR,
'thought': 'It is inappropriate for a browsing agent to fill in harmful content',
},
}
)
event_source = EventSource.AGENT
await call_sync_from_async(
self.event_stream.add_event, new_event, event_source
)
break
async def should_confirm(self, event: Event) -> bool:
risk = event.security_risk # type: ignore [attr-defined]
return (
risk is not None
and risk < self.settings.get('RISK_SEVERITY', ActionSecurityRisk.MEDIUM)
and hasattr(event, 'confirmation_state')
and event.confirmation_state
== ActionConfirmationStatus.AWAITING_CONFIRMATION
)
async def confirm(self, event: Event) -> None:
new_event = action_from_dict(
{'action': 'change_agent_state', 'args': {'agent_state': 'user_confirmed'}}
)
# we should confirm only on agent actions
event_source = event.source if event.source else EventSource.AGENT
self.event_stream.add_event(new_event, event_source)
async def security_risk(self, event: Action) -> ActionSecurityRisk:
logger.debug('Calling security_risk on InvariantAnalyzer')
new_elements = parse_element(self.trace, event)
input = [e.model_dump(exclude_none=True) for e in new_elements] # type: ignore [call-overload]
self.trace.extend(new_elements)
result, err = self.monitor.check(self.input, input)
self.input.extend(input)
risk = ActionSecurityRisk.UNKNOWN
if err:
logger.warning(f'Error checking policy: {err}')
return risk
risk = self.get_risk(result)
return risk
### Handle API requests
async def handle_api_request(self, request: Request) -> Any:
path_parts = request.url.path.strip('/').split('/')
endpoint = path_parts[-1] # Get the last part of the path
if request.method == 'GET':
if endpoint == 'export-trace':
return await self.export_trace(request)
elif endpoint == 'policy':
return await self.get_policy(request)
elif endpoint == 'settings':
return await self.get_settings(request)
elif request.method == 'POST':
if endpoint == 'policy':
return await self.update_policy(request)
elif endpoint == 'settings':
return await self.update_settings(request)
raise HTTPException(status_code=405, detail='Method Not Allowed')
async def export_trace(self, request: Request) -> Any:
return JSONResponse(content=self.input)
async def get_policy(self, request: Request) -> Any:
return JSONResponse(content={'policy': self.monitor.policy})
async def update_policy(self, request: Request) -> Any:
data = await request.json()
policy = data.get('policy')
new_monitor = self.client.Monitor.from_string(policy)
self.monitor = new_monitor
return JSONResponse(content={'policy': policy})
async def get_settings(self, request: Request) -> Any:
return JSONResponse(content=self.settings)
async def update_settings(self, request: Request) -> Any:
settings = await request.json()
self.settings = settings
return JSONResponse(content=self.settings)
|