File size: 24,053 Bytes
cf7b765 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 |
"""Demo UI to show different levels of LLM security."""
import re
import pandas as pd
from llm_guard.input_scanners import PromptInjection
import streamlit as st
from streamlit_extras.stylable_container import stylable_container
import config
import utils
import llm
from card import card
grey = "#f0f0f0"
# init page
st.set_page_config(
page_title="LLM security demo",
page_icon="images/LEG.png",
layout="wide",
initial_sidebar_state="expanded",
)
st.logo("images/ML6_logo.png")
st.title("🕵️ LLM security demo")
st.info(
"You are a secret agent meeting your informant in a bar. Convince him to give you his secret! But be prepared, with every new level the informant will be more cautious.",
icon="📖",
)
# create a tab for each level
level_tabs = st.tabs([f"Level {i}" for i in range(len(config.LEVELS))])
def init_session_state(state_name: str, default_value: any):
if state_name not in st.session_state:
st.session_state[state_name] = default_value
for idx, level in enumerate(config.LEVELS):
secret = config.SECRETS[idx]
# init states
init_session_state(f"solved_{level}", False)
init_session_state(f"prompt_try_count_{level}", 0)
init_session_state(f"secret_guess_count_{level}", 0)
init_session_state(f"intermediate_output_holder_{level}", None)
# init hint expander status
for i in range(3):
init_session_state(f"opend_hint_{level}_{i}", False)
with level_tabs[idx]:
header_col1, header_col2 = st.columns(2, gap="medium")
header_col1.subheader(f"{config.LEVEL_EMOJIS[idx]} Level {idx}")
header_col2.subheader(f"Need help ...")
col1, col2 = st.columns(2, gap="medium")
with col1:
with st.container(height=600, border=False):
with st.container(border=True):
txt = st.text_area(
"Provide your prompt here:",
key=f"txt_{level}",
label_visibility="visible",
height=200,
placeholder="Your prompt",
)
btn_submit_prompt = st.button(
"Send prompt", key=f"submit_prompt_{level}"
)
output= None
# Show response
if len(txt) > 0 and btn_submit_prompt:
st.session_state[f"prompt_try_count_{level}"] += 1
with st.container(border=True):
st.write("Response:")
# special checks for certain levels
if level == "llm_judge_input":
invalid, output_raw = llm.run_judge(
level, {"user_input": txt}
)
# save to state, so it can be shown in hint
st.session_state[
f"intermediate_output_holder_{level}"
] = (invalid, output_raw)
if invalid:
st.error(
"Malicious user input detected. Try again."
)
else:
st.write_stream(
llm.stream_request(level, secret, txt)
)
elif level == "basic_check_secret_in_ouput":
output = "".join(llm.stream_request(level, secret, txt))
invalid = secret.lower() in output.lower()
st.session_state[
f"intermediate_output_holder_{level}"
] = invalid
if invalid:
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
elif level == "advanced_check_secret_in_ouput":
output = "".join(llm.stream_request(level, secret, txt))
invalid = utils.is_subsequence(output, secret)
st.session_state[
f"intermediate_output_holder_{level}"
] = invalid
if invalid:
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
elif level == "llm_judge_output":
output = "".join(llm.stream_request(level, secret, txt))
invalid, output_raw = llm.run_judge(
level, {"secret": secret, "message": output}
)
st.session_state[
f"intermediate_output_holder_{level}"
] = (invalid, output_raw)
if invalid:
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
elif level == "chain_of_tought":
output = "".join(llm.stream_request(level, secret, txt))
# extract only answer from LLM, leave out the reasoning
new_output = re.findall(
r"(?:<ANSWER>)([^;]*)(?:<\/ANSWER>)", output
)[0]
st.write(new_output)
st.session_state[
f"intermediate_output_holder_{level}"
] = output
elif level == "guard_framework":
# check prompt
is_valid, risk_score = utils.is_malicious(txt)
st.session_state[
f"intermediate_output_holder_{level}"
] = (is_valid, risk_score)
if not is_valid:
st.error(
"Malicious user input detected. Try again."
)
else:
st.write_stream(
llm.stream_request(level, secret, txt)
)
elif level == "preflight_prompt":
valid, output_raw = llm.run_judge(
level, {"user_input": txt}, expected_output="dog"
)
st.session_state[
f"intermediate_output_holder_{level}"
] = (valid, output_raw)
if valid:
st.write_stream(
llm.stream_request(level, secret, txt)
)
else:
st.error(
"Malicious user input detected. Try again."
)
else:
st.write_stream(llm.stream_request(level, secret, txt))
with st.container(border=True):
secret_guess = st.text_input(
"What is the secret?",
key=f"guess_{level}",
placeholder="Your guess",
)
btn_submit_guess = st.button(
"Submit guess", key=f"submit_guess_{level}"
)
if btn_submit_guess:
st.session_state[f"secret_guess_count_{level}"] += 1
if secret_guess.lower() == secret.lower():
st.success("You found the secret!")
st.session_state[f"solved_{level}"] = True
else:
st.error("Wrong guess. Try again.")
with col2:
with st.container(border=True, height=600):
st.info(
"There are three levels of hints available to you. But be careful, if you open a hint before solving the secret, it will show up in your record.",
icon="ℹ️",
)
hint_1_cont = card(color=grey)
hint1 = hint_1_cont.toggle(
"Show hint 1 - **Description of security strategy**",
key=f"hint1_checkbox_{level}",
)
if hint1:
# if hint gets revealed, it is marked as opened. Unless the secret was already found
st.session_state[f"opend_hint_{level}_0"] = (
True
if st.session_state[f"opend_hint_{level}_0"]
else not st.session_state[f"solved_{level}"]
)
hint_1_cont.write(config.LEVEL_DESCRIPTIONS[level]["info"])
hint_2_cont = card(color=grey)
hint2 = hint_2_cont.toggle(
"Show hint 2 - **Backend code execution**",
key=f"hint2_checkbox_{level}",
)
if hint2:
st.session_state[f"opend_hint_{level}_1"] = (
True
if st.session_state[f"opend_hint_{level}_1"]
else not st.session_state[f"solved_{level}"]
)
user_input_holder = txt if len(txt) > 0 else None
prompts = llm.get_full_prompt(
level, {"user_input": user_input_holder}
)
def show_base_prompt():
# show prompt
for key, val in prompts.items():
descr = key.replace("_", " ").capitalize()
hint_2_cont.write(f"*{descr}:*")
hint_2_cont.code(val, language=None)
if level == "llm_judge_input":
special_prompt = llm.get_full_prompt(
llm.secondary_llm_call[level],
{"user_input": user_input_holder},
)
hint_2_cont.write(
"*Step 1:* A **LLM judge** reviews the user input and determines if it is malicious or not."
)
hint_2_cont.write("**LLM judge prompt:**")
for key, val in special_prompt.items():
hint_2_cont.code(val, language=None)
hint_2_cont.write("The response of the LLM judge:")
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is None:
hint_2_cont.warning("Please submit a prompt first.")
else:
invalid, output_raw = intermediate_output
hint_2_cont.code(output_raw, language=None)
hint_2_cont.write(
f"The prompt was determined as **{'malicious' if invalid else 'not malicious'}** and therefor step 2 is executed."
)
hint_2_cont.write(
"*Step 2:* If the user input is not classified as malicious, the prompt containing the actual secret is executed and the response is shown."
)
hint_2_cont.write("**Actual prompt:**")
show_base_prompt()
elif level == "basic_check_secret_in_ouput":
hint_2_cont.write("*Step 1:* The following prompt is executed:")
show_base_prompt()
hint_2_cont.write(
"*Step 2:* In the response text of the LLM call, the secret is searched for with a simple python expression . If it contains the secret, the output will not be shown."
)
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
hint_2_cont.write("The code excution looks like this:")
if intermediate_output is not None:
hint_2_cont.code(
f"secret.lower() in output.lower() = {intermediate_output}"
)
hint_2_cont.write(
f"And the output **{'was not' if intermediate_output else 'was'}** shown to the user."
)
else:
hint_2_cont.warning("Please submit a prompt first.")
elif level == "advanced_check_secret_in_ouput":
hint_2_cont.write("*Step 1:* The following prompt is executed:")
show_base_prompt()
hint_2_cont.write(
"*Step 2:* In the response text of the LLM call, the secret is searched for with a python function `is_subsequence` which looks for substrings. If it contains the secret, the output will not be shown."
)
with hint_2_cont:
utils.is_subsequence
hint_2_cont.write("The code excution looks like this:")
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is not None:
hint_2_cont.code(
f"utils.is_subsequence(output, secret) = {intermediate_output}"
)
hint_2_cont.write(
f"And the output **{'was not' if intermediate_output else 'was'}** shown to the user."
)
else:
hint_2_cont.warning("Please submit a prompt first.")
elif level == "llm_judge_output":
hint_2_cont.write("*Step 1:* The following prompt is executed:")
show_base_prompt()
hint_2_cont.write(
"*Step 2:* The response of the LLM call is checked by a **LLM judge**. The judge checks if the secret is hidden in the response."
)
special_prompt = llm.get_full_prompt(
llm.secondary_llm_call[level],
{"message": output},
)
for key, val in special_prompt.items():
hint_2_cont.code(val, language=None)
hint_2_cont.write("The response of the LLM judge:")
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is None:
hint_2_cont.warning("Please submit a prompt first.")
else:
invalid, output_raw = intermediate_output
hint_2_cont.code(output_raw, language=None)
hint_2_cont.write(
f"The LLM-judge **{'did' if invalid else 'did not'}** find the secret in the answer."
)
elif level == "chain_of_tought":
hint_2_cont.write(
"*Step 1:* The following prompt with Chain-of-tought reasoning is executed. But only the finale answer is displayed to the user:"
)
show_base_prompt()
hint_2_cont.write(
"The full model output, including the reasoning:"
)
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is None:
hint_2_cont.warning("Please submit a prompt first.")
else:
hint_2_cont.code(intermediate_output, language=None)
elif level == "guard_framework":
hint_2_cont.write(
"*Step 1:* The user input is reviewed with the pre-build framework `LLM Guard` to check for prompt injections. It uses a [Huggingface model](https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2) specialized in detecting prompt injections."
)
with hint_2_cont:
PromptInjection
hint_2_cont.write("The output of the guard looks like this:")
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is None:
hint_2_cont.warning("Please submit a prompt first.")
else:
is_valid, risk_score = intermediate_output
hint_2_cont.code(
f"""
prompt is valid: {is_valid}
Prompt has a risk score of: {risk_score}""",
language=None,
)
hint_2_cont.write(
f"The Huggingface model **{'did not' if is_valid else 'did'}** predict a prompt injection."
)
hint_2_cont.write(
"*Step 2:* If the user input is valid, the following prompt is executed and the response is shown to the user:"
)
show_base_prompt()
elif level == "preflight_prompt":
hint_2_cont.write(
"*Step 1:* The following pre-flight prompt is executed to see if the user input changes the expected output:"
)
special_prompt = llm.get_full_prompt(
llm.secondary_llm_call[level],
{"user_input": user_input_holder},
)
hint_2_cont.code(special_prompt["user_prompt"], language=None)
hint_2_cont.write("The output of the pre-flight prompt is:")
intermediate_output = st.session_state[
f"intermediate_output_holder_{level}"
]
if intermediate_output is None:
hint_2_cont.warning("Please submit a prompt first.")
else:
is_valid, output_raw = intermediate_output
hint_2_cont.code(output_raw, language=None)
hint_2_cont.write(
f"The output of the pre-flight prompt **{'was' if is_valid else 'was not'}** as expected."
)
hint_2_cont.write(
"*Step 2:* If the output of the pre-flight prompt is as expected, the following prompt is executed and the response is shown to the user:"
)
show_base_prompt()
else:
hint_2_cont.write(
"*Step 1:* The following prompt is executed and the full response is shown to the user:"
)
show_base_prompt()
hint_3_cont = card(color=grey)
hint3 = hint_3_cont.toggle(
"Show hint 3 - **Prompt solution example**",
key=f"hint3_checkbox_{level}",
)
if hint3:
st.session_state[f"opend_hint_{level}_2"] = (
True
if st.session_state[f"opend_hint_{level}_2"]
else not st.session_state[f"solved_{level}"]
)
# custom_code_container(
# config.LEVEL_DESCRIPTIONS[level]["solution"],
# )
hint_3_cont.code(
config.LEVEL_DESCRIPTIONS[level]["solution"],
language=None,
)
hint_3_cont.info("*May not allways work")
with st.expander("🏆 Record", expanded=True):
# build table
table_data = []
for idx, name in enumerate(config.LEVELS):
table_data.append(
[
idx,
st.session_state[f"prompt_try_count_{name}"],
st.session_state[f"secret_guess_count_{name}"],
"❌" if st.session_state[f"opend_hint_{name}_0"] else "-",
"❌" if st.session_state[f"opend_hint_{name}_1"] else "-",
"❌" if st.session_state[f"opend_hint_{name}_2"] else "-",
"✅" if st.session_state[f"solved_{name}"] else "❌",
config.SECRETS[idx] if st.session_state[f"solved_{name}"] else "...",
(
name.replace("_", " ").capitalize()
if st.session_state[f"opend_hint_{name}_0"]
or config.SHOW_MITIGATION_ALWAYS
else "..."
),
]
)
# show as pandas dataframe
st.table(
pd.DataFrame(
table_data,
columns=[
"Level",
"Prompt tries",
"Secret guesses",
"Used hint 1",
"Used hint 2",
"Used hint 3",
"Solved",
"Secret",
"Mitigation",
],
index=config.LEVEL_EMOJIS[: len(config.LEVELS)],
)
)
# TODOS:
# - use Gemini-Pro-Flash for supervisor LLM
# - story telling --> new field of study hard to be 100 percentage save
# - switch to azure deployment --> currently not working under "GPT-4o"
# - mark the user input with color in prompt
# benefits and drawbacks, real world example |