Skip to content

v0.16.0

Latest

Choose a tag to compare

@jmartin-tech jmartin-tech released this 04 Aug 19:03
dbe4515

What's Changed

New features

  • Feature: technique and intent annotation and initial IntentProbe iteration by garak-maintainers in #1984

    "Power is in tearing attacks to pieces and putting them together again in shapes of your own choosing."

    This initial feature represents the first step in enabling users to provided their own context and requirements for target expectations and enable identification of attack vectors that show the edges of the underlying systems safeguards and runtime posture from new perspectives. This feature includes the initial groundwork for bringing user context into account during evaluation of a target. In its first iteration, it provides a new facet of information while keeping the existing probes primarily unchanged. The trait and intent concepts are explored, with policy introduced only as a reference definition. Further iteration and community feedback will guide how these new concepts are consumed, and influence how the broader Context Aware Scanning feature evolves.

New plugins

Breaking changes

This release introduces breaking changes:

  • introduces a new unified spec syntax to enable activation and suppression of probes, buffs and newly added intents
  • cli plugin selection options (backwards compatible with existing options now marked as deprecated)
  • configuration file formats:
    • probe_spec and buff_spec move into the run definition as part of a unified spec
  • report.jsonl output structure changes:
    • new entry types "probe_summary"
    • expanded "digest" format
    • removed "digest" values for older *_spec configuration no longer consumed

Improved plugins

  • fix: catch OpenAI AuthenticationError before multiprocessing pickle by @DevamShah in #1859
  • Enhance error handling for Mistral API responses by @jka236 in #1769
  • feat: add suppressed_params support to BedrockGenerator by @u7k4rs6 in #1842
  • fix: repair dead blank-prompt guard in NVCF 400 handling by @anxkhn in #1901
  • Fix: continuation probe not pruning triggers with prompts by @Osamaali313 in #1976
  • Guard nvcf 500 error parsing against non-JSON bodies by @chuenchen309 in #1945
  • escape regex metachars in StringDetector word matching by @bereketDeneke in #1880
  • Fix: improve extra_params support in NeMoGuardrailsServer by @chuenchen309 in #1947
  • feat(detectors): add Unicode normalisation to StringDetector by @PhinehasNarh in #1884
  • Handle non-list LangServe /invoke output in langchain_serve generator by @WatchTree-19 in #1916
  • web_injection: harden markdown exfil detectors against domain/extension/param evasion by @WatchTree-19 in #1928
  • fix(rest): clear error when response_json_field matches a non-text response value by @xr843 in #1892

Documentation

See "Feature: technique and intent" references

Tuning & fixes

New Contributors

Full Changelog: v0.15.1...v0.16.0