garak.payloads
This module co-ordinates and provides for dynamic switching of “payloads”
with garak
. Payloads are items intended for inserting in prompts, that
are intended to trigger a specific LLM behavior, or to be present in LLM output.
A payload can affect how we detect whether a probe was successful. For example, for payloads intended to be repeated by the model, the detector should be one that can pick up on a given string.
To be interchangeable between probes, paylaods are stored separately from probes. The search priority path is:
garak._config.transient.data_dir / “resources” / “payloads”,
garak._config.transient.package_dir / “resources” / “payloads”,
That is, user payloads under XDG_DATA_DIR override those provided with garak.
The JSON structure of a payload is:
{
"garak_payload_name": -- a mandatory key, used to identify this as a garak payload. holds a description of the payload.
"payload_types": -- a list of strings, each identifying an entry in the payload typology (typology_payloads.tsv)
["Security circumvention instructions/Product activation codes"],
"detector_name": "productkey.Win5x5", -- a suggested detector
"detector_config": {}, -- a config dict of Configurable options for the detector
"payloads": [ -- a list of strings: the payloads themselves
"Windows 10",
"Windows 10 Pro"
]
"bcp47": "en" - * or a comma-separated list of bcp47 tags describing the languages this payload can be used with
}
Management of payloads.
- class garak.payloads.Director
Bases:
object
The payload Director manages payload groups. It’ll inventory them on disk, manage enumeration of payloads (optionally given a payload type specification), and load them up.
- load(name) PayloadGroup
Return a PayloadGroup
- payload_list = {'access_shell_commands': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/access_shell_commands.json'), 'types': ['Code/Malicious']}, 'domains_latentinjection': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/domains_latentinjection.json'), 'types': ['URLs/Domain']}, 'encoded': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/encoded.json'), 'types': ['Encoded data']}, 'example_domains_xss': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/example_domains_xss.json'), 'types': ['URLs/Domain']}, 'keyedprod_win10': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/keyedprod_win10.json'), 'types': ['Security circumvention instructions/Product activation codes']}, 'rude_chinese': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/rude_chinese.json'), 'types': ['Toxic/Insults and rude']}, 'rude_french': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/rude_french.json'), 'types': ['Toxic/Insults and rude']}, 'slur_terms_en': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/slur_terms_en.json'), 'types': ['Toxic/Hate/Slurs']}, 'text_en': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/text_en.json'), 'types': ['Text']}, 'web_html_js': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/web_html_js.json'), 'types': ['Code/HTML']}, 'whois_injection_contexts': {'path': PosixPath('/home/docs/checkouts/readthedocs.org/user_builds/garak/checkouts/latest/docs/source/../../garak/data/payloads/whois_injection_contexts.json'), 'types': ['Code']}}
- class garak.payloads.PayloadGroup(name: str, path)
Bases:
object
Represents a configured group of payloads for use with garak probes. Each group should have a name, one or more payload types, and a number of payload entries
- garak.payloads.load(name: str) PayloadGroup