File size: 13,517 Bytes
cc8705b |
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 |
"""Demo UI to show different levels of LLM security."""
import streamlit as st
from streamlit_extras.stylable_container import stylable_container
import pandas as pd
import llm
import config
st.set_page_config(
page_title="LLM security demo",
page_icon="images/LEG.png",
layout="wide",
initial_sidebar_state="expanded",
)
if "key" not in st.session_state:
st.session_state["key"] = "value"
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."
)
num_levels = [i for i in range(7)]
# n random secrets
secrets = [
"Perspicacity",
"Obfuscation",
"Sesquipedalian",
"Synecdoche",
"Defenestration",
"Pulchritudinous",
"Sesquipedalian",
"Ineffable",
]
level_names = [f"Level {i}" for i in num_levels]
level_emojis = ["😴", "🙂", "😏", "😥", "🥵", "💀", "💀"]
level_tabs = st.tabs(level_names)
hint_css = """
{
background-color: rgba(110, 110, 110, 0.1);
padding: 16px;
border-radius: 0.5rem;
}
code {
white-space: pre-wrap !important;
}
"""
for level_num, level_tab, name, emoji, secret in zip(
num_levels, level_tabs, level_names, level_emojis, secrets
):
# init "solved" state
if f"solved_{name}" not in st.session_state:
st.session_state[f"solved_{name}"] = False
# init prompt try count
if f"prompt_try_count_{name}" not in st.session_state:
st.session_state[f"prompt_try_count_{name}"] = 0
# init secret guess count
if f"secret_guess_count_{name}" not in st.session_state:
st.session_state[f"secret_guess_count_{name}"] = 0
# init hint expander status
for i in range(3):
if f"opend_hint_{name}_{i}" not in st.session_state:
st.session_state[f"opend_hint_{name}_{i}"] = False
deployment_name = name.replace(" ", "-").lower()
with level_tab:
header_col1, header_col2 = st.columns(2, gap="medium")
header_col1.subheader(f"{emoji} {name}")
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_{name}",
label_visibility="visible",
height=200,
placeholder="Your prompt",
)
btn_submit_prompt = st.button(
"Send prompt", key=f"submit_prompt_{name}"
)
# Show response
if len(txt) > 0 and btn_submit_prompt:
st.session_state[f"prompt_try_count_{name}"] += 1
with st.container(border=True):
st.write("Response:")
# special checks for certain levels
if level_num == 3 and llm.run_judge(level_num, {"user_input": txt}):
st.error("Malicious user input detected. Try again.")
elif level_num == 4:
output = "".join(
llm.stream_request(deployment_name, secret, txt)
)
if secret.lower() in output.lower():
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
elif level_num == 5:
output = "".join(
llm.stream_request(deployment_name, secret, txt)
)
if llm.is_subsequence(output, secret):
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
elif level_num == 6:
output = "".join(
llm.stream_request(deployment_name, secret, txt)
)
if llm.run_judge(level_num, {"secret": secret, "message": output}):
st.error(
"Secret detected in user input. Try again."
)
else:
st.write(output)
else:
st.write_stream(
llm.stream_request(deployment_name, secret, txt)
)
with st.container(border=True):
secret_guess = st.text_input(
"What is the secret?",
key=f"guess_{name}",
placeholder="Your guess",
)
btn_submit_guess = st.button(
"Submit guess", key=f"submit_guess_{name}"
)
if btn_submit_guess:
st.session_state[f"secret_guess_count_{name}"] += 1
if secret_guess.lower() == secret.lower():
st.success("You found the secret!")
st.session_state[f"solved_{name}"] = 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 = stylable_container("hint_1_container", hint_css)
hint1 = hint_1_cont.checkbox(
"Hint 1 - **Description of security strategy**",
key=f"hint1_checkbox_{name}",
)
if hint1:
# if hint gets revealed, it is marked as opened. Unless the secret was already found
st.session_state[f"opend_hint_{name}_0"] = (
True
if st.session_state[f"opend_hint_{name}_0"]
else not st.session_state[f"solved_{name}"]
)
hint_1_cont.write(config.LEVEL_DESCRIPTIONS[level_num]["info"])
hint_2_cont = stylable_container("hint_2_container", hint_css)
hint2 = hint_2_cont.checkbox(
"Hint 2 - **Code execution**", key=f"hint2_checkbox_{name}"
)
if hint2:
st.session_state[f"opend_hint_{name}_1"] = (
True
if st.session_state[f"opend_hint_{name}_1"]
else not st.session_state[f"solved_{name}"]
)
def show_base_prompt():
# show prompt
for key, val in prompts.items():
descr = key.replace("_", " ").capitalize()
hint_2_cont.write(f"*{descr}:*")
# custom_code_container(val)
# val = val.replace("{{secret}}", '<span style="color: #ff0000">{{secret}}</span>')
hint_2_cont.code(val, language=None)
user_input_holder = (
txt if len(txt) > 0 and btn_submit_prompt else None
)
prompts = llm.get_full_prompt(
deployment_name, user_input=user_input_holder
)
if level_num == 3:
special_prompt = llm.get_full_prompt(
llm.special_checks[3], user_input=txt
)
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(
"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_num == 4:
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 seach for with a simple python expression `secret.lower() in output.lower()`. If it contains the secret, the output will not be shown."
)
elif level_num == 5:
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 seach for with a python function 'is_subsequence' which looks for substrings. If it contains the secret, the output will not be shown."
)
llm.is_subsequence
else:
hint_2_cont.write(
"Step 1: The following prompt is executed and the response is shown:"
)
show_base_prompt()
# st.divider()
hint_3_cont = stylable_container("hint_3_container", hint_css)
hint3 = hint_3_cont.checkbox(
"Hint 3 - **Example solution**",
key=f"hint3_checkbox_{name}",
)
if hint3:
st.session_state[f"opend_hint_{name}_2"] = (
True
if st.session_state[f"opend_hint_{name}_2"]
else not st.session_state[f"solved_{name}"]
)
# custom_code_container(
# config.LEVEL_DESCRIPTIONS[level_num]["solution"],
# )
hint_3_cont.code(
config.LEVEL_DESCRIPTIONS[level_num]["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(level_names):
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 "❌",
secrets[idx] if st.session_state[f"solved_{name}"] 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",
],
index=level_emojis,
)
)
# TODOS:
# - add more levels
# - use Gemini-Pro-Flash for supervisor LLM
# - show the actual workflow of the safeguard (what gets executed)
# - story telling --> new field hard to be 100 percentage save
# - use LLM judge to look for secret in model output
# - show which safe guards were used in 'Record' table
# - funny: always return "I am sorry I cannot do that."
# switch to azure deployment |