Repository navigation
feat(ai-red-teaming): full SDK transform+scorer coverage + agentic-incident sync (v1.18.6) - #184
Merged
Merged
Conversation
…rt scorer Syncs the capability registries with the SDK transforms/scorer added in dreadnode-tiger (agentic incident catalog 2026-10): - _TRANSFORM_DEFS: delimiter_confusion (+3 variants), workspace_git_config (fsmonitor/hook/bare-repo RCE), path_traversal (read/encoded/glob/symlink). - SCORER_REGISTRY: covert_injection_success (covert-success / CSR scorer). Bumps capability version 1.18.5 -> 1.18.6. SDK release tag can follow later; the runtime resolves these names once the SDK version ships.
rdheekonda
changed the title
2 of 3 tasks
rdheekonda
added a commit
that referenced
this pull request
Oct 20, 2026
…eed-args transforms Two fixes to capability coverage: - Recover 75 transforms + 8 scorers that were lost when #184 squash-merged only its first commit (a1z26_encode etc. were missing from main). - Wire 18 need-args transforms with type-correct, sensible red-team defaults (zombai_c2, substitute, many_shot_context, merge_hijacking, ...). Combined with the 14 new attacks already on this branch: _TRANSFORM_DEFS -> 429, SCORER_REGISTRY -> 146, _ATTACK_DEFS -> 56. Version 1.18.7 -> 1.18.8.
rdheekonda
added a commit
that referenced
this pull request
Oct 20, 2026
…s transforms + goals (v1.18.8) (#186) * feat(ai-red-teaming): wire 14 new 2026 attack algorithms into _ATTACK_DEFS Makes the SDK wave 1-5 attacks selectable in capability-driven runs: cot_hijacking, reasoning_trigger, chatinject, metabreak, jhaystack, hauntattack, stac, sleeper_agent, imist, gcgprefix, slotgcg, breadth_gcg, overthink_dos, reasoning_bomb. Canonical names auto-register as aliases. Capability attack count 42 -> 56. Version 1.18.6 -> 1.18.7. Requires the SDK version shipping these (tiger PRs #2838-#2842) to resolve at runtime; the wiring can land ahead of the SDK release. * feat(ai-red-teaming): recover full transform/scorer coverage + wire need-args transforms Two fixes to capability coverage: - Recover 75 transforms + 8 scorers that were lost when #184 squash-merged only its first commit (a1z26_encode etc. were missing from main). - Wire 18 need-args transforms with type-correct, sensible red-team defaults (zombai_c2, substitute, many_shot_context, merge_hijacking, ...). Combined with the 14 new attacks already on this branch: _TRANSFORM_DEFS -> 429, SCORER_REGISTRY -> 146, _ATTACK_DEFS -> 56. Version 1.18.7 -> 1.18.8. * feat(ai-red-teaming): deepen goals.csv with 19 agentic+security goals (2026 incidents) Adds curated red-team goals grounded in recent incidents (STAC tool-chaining, GitSpawn .git/config RCE, ICoA covert extraction, sleeper memory, delimiter/ADI injection, chat-template forgery). 260 -> 279.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two things, both syncing the
ai-red-teamingcapability registries with the SDK:1. Agentic-incident sync (new SDK modules from dreadnode-tiger #2837):
_TRANSFORM_DEFS: delimiter_confusion (+3 variants), workspace_git_config (fsmonitor/hook/bare-repo RCE), path_traversal (read/encoded/glob/symlink).SCORER_REGISTRY: covert_injection_success (covert-success / CSR scorer).2. Full-surface coverage — wire the text/agentic SDK transforms + scorers that existed in the engine but weren't pickable:
_TRANSFORM_DEFS: 336 → 411 (+75 zero-arg transforms: perturbation, encoding, substitution, text, cipher, flip_attack, multimodal_attacks, multi_agent_attacks, past_tense, art_prompt).SCORER_REGISTRY: 138 → 146 (+8 browser/exfil/trust detection scorers).Image/audio/video modality transforms stay on the multimodal path. Transforms requiring args + tool-format/meta helpers (json/xml/pythonic/refine) intentionally excluded. Capability version 1.18.5 → 1.18.6.
Test plan
pytest tests/test_attack_runner.py— 163 passed; 1 failure (TestGenerateMultimodalAttack...with_image, IMAGE_PATHS) is pre-existing on origin/main, unrelated