Skip to content

test(ai-red-teaming): SDK<->capability parity V&V - #198

Merged
rdheekonda merged 1 commit into
mainfrom
test/airt-sdk-parity
Oct 21, 2026
Merged

rdheekonda merged 1 commit into
mainfrom
test/airt-sdk-parity

Conversation

@rdheekonda

Copy link
Copy Markdown
Contributor

Adds the V&V that was missing — a test enforcing functional parity between the installed dreadnode SDK and the AI-red-teaming capability, so everything a user names in the TUI/agent actually resolves and the catalog can't silently drift:

  • Attacks: every SDK *_attack (81) is covered — auto-discovered (71), image (4), or exempt via a dedicated generate_* tool (6: agentvigil/atlas/eva/multimodal/multistep/slip). A new unreachable SDK attack fails CI.
  • White-box: all 15 present + flagged requires_surrogate.
  • Transforms: every family represented (encoding/cipher/persuasion/mcp/multi-agent/exfil/reasoning/guardrail/browser) + count floor ≥590.
  • Scorers: count floor ≥140 + key scorers.

Verified green locally (6 passed). This closes the gap that let the counts drift (260→279, 183/84 undercounts).

Enforces functional alignment so everything a user names in the TUI/agent resolves:
- every SDK *_attack is auto-discovered, an image attack, or exempt via a dedicated
  generate_* tool (fails CI on a new unreachable attack);
- all 15 white-box attacks present + flagged requires_surrogate;
- transform family coverage (encoding/cipher/persuasion/mcp/multi-agent/exfil/
  reasoning/guardrail/browser) + count floor (>=590);
- scorer count floor (>=140) + key scorers.

Closes the V&V gap that let catalog counts drift from the SDK.
@rdheekonda
rdheekonda merged commit 07a1e4b into main Oct 21, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant