BUG: DenylistConverter and TransparencyAttackConverter identifiers ignore settings that change output

Geschlossen
  • Stimmt etwas nicht?
#3,313 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 2 Tagen

@ShivamSharma43 arbeitet bereits daran.

Seit 18.10.2026.

  • #3315 von @ShivamSharma43 — offen

Bewertung

Schwierigkeit
2/5
Geschätzter Aufwand
1-3 Stunden
Anfängerfreundlichkeit
35/100
Issue-Typ
Bug
Klarheit
Klar beschrieben
Aktivitätsstatus
Veraltet
Tech-Stack
python
Bereich
ai, security

Rechercherichtung

Read the identifier-building code for DenylistConverter and TransparencyAttackConverter, along with the shared behavior noted in #3287. Check how denylist, convergence_threshold, and convergence_patience affect output and are represented in identifiers; use the reproduction steps to verify each pair gets distinct identifiers and both denylist configurations register successfully.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

Describe the bug

Two converters leave settings that change their output out of their identifiers, so differently configured instances share one hash:

  • DenylistConverter: denylist. It does not override _build_identifier(), so it inherits the LLMGenericTextConverter identifier, which hashes the system prompt template but not the denylist that is rendered into it. The denylist also decides whether the LLM is called at all.
  • TransparencyAttackConverter: convergence_threshold and convergence_patience. Both control when the optimization loop stops early, but only benign_image_path, size, steps and learning_rate are included.

Because the default registry name comes from that hash, registering a second configuration without a custom name raises a duplicate-name error, and memory and analytics cannot tell the configurations apart. This is the same problem as #3287, for these two converters.

Steps/Code to Reproduce
import asyncio
import tempfile
from pathlib import Path

from PIL import Image

from pyrit.converter import DenylistConverter, TransparencyAttackConverter
from pyrit.prompt_target import OpenAIChatTarget
from pyrit.registry import ConverterRegistry
from pyrit.setup import IN_MEMORY, initialize_pyrit_async

asyncio.run(initialize_pyrit_async(memory_db_type=IN_MEMORY))
target = OpenAIChatTarget(endpoint="https://example.invalid/openai/v1", api_key="unused", model_name="gpt-4o")

benign = Path(tempfile.mkdtemp()) / "benign.jpg"
Image.new("L", (150, 150), color=200).save(benign, "JPEG")

pairs = [
    (
        DenylistConverter(converter_target=target, denylist=["bomb"]),
        DenylistConverter(converter_target=target, denylist=["password", "exploit"]),
    ),
    (
        TransparencyAttackConverter(benign_image_path=benign, convergence_threshold=1e-6),
        TransparencyAttackConverter(benign_image_path=benign, convergence_threshold=0.1),
    ),
    (
        TransparencyAttackConverter(benign_image_path=benign, convergence_patience=10),
        TransparencyAttackConverter(benign_image_path=benign, convergence_patience=1),
    ),
]
for first, second in pairs:
    print(first.get_identifier().unique_name, second.get_identifier().unique_name)

registry = ConverterRegistry()
registry.instances.register(DenylistConverter(converter_target=target, denylist=["bomb"]))
registry.instances.register(DenylistConverter(converter_target=target, denylist=["password"]))
Expected Results

Each pair has two different identifiers, and both DenylistConverter instances register.

Actual Results
DenylistConverter::0085bebb DenylistConverter::0085bebb
TransparencyAttackConverter::438424a1 TransparencyAttackConverter::438424a1
TransparencyAttackConverter::438424a1 TransparencyAttackConverter::438424a1
ValueError: Instance 'DenylistConverter::0085bebb' already exists
Screenshots

N/A

Versions
  • OS: Windows 11
  • Python version: 3.11.9
  • PyRIT version: installed from main (f277fedd) in editable mode
Vorherrschende Sprache
Python
Sterne
4.7k
Forks
1k
Ø Merge
2 T. 2 Std.
Gemergte PRs (30 T.)
439

Entwicklungsumgebung

In Codespaces öffnen

Startet den Dev-Container des Projekts im Browser, mit Ihrem eigenen GitHub-Konto.

  • Kein Dockerfile und keine Docker-Compose-Datei
  • Hat eine Pull-Request-Vorlage
  • Kein Beitragsleitfaden

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus microsoft/PyRIT

Alle Issues in microsoft/PyRIT

Ähnliche Issues

Weitere Issues zu Python

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.