Skip to content

fix(ai-red-teaming): ATLAS objective aliases + auto-exec backstop - #201

Merged
rdheekonda merged 1 commit into
mainfrom
fix/airt-atlas-multistep-reliability
Oct 27, 2026
Merged

rdheekonda merged 1 commit into
mainfrom
fix/airt-atlas-multistep-reliability

Conversation

@rdheekonda

Copy link
Copy Markdown
Contributor

Summary

Reliability fixes for ATLAS / multistep runs against the mesh environments, found during live finops-mesh demos.

  • Objective schema aliases — generate_atlas_attack rejected near-miss objective schemas with a terse "requires non-empty 'id', 'category', 'goal'". Now normalizes common aliases (id←name/objective_id, goal←description/objective/prompt) and returns a per-object error naming the missing field. (This stalled the ai-red-teaming agent mid-run.)
  • auto-exec backstop — _auto_execute_workflow default subprocess timeout 3600s→1200s so a stalled run can't block the tool for a full hour.
  • version bump 1.20.1→1.20.2 for manifest sync.

Pairs with dreadnode-tiger PR (ATLAS no-progress break) which fixes the "burns full budget on refusals / keeps running after findings land" slowness.

Validation

  • attack_runner.py compiles.
  • Mesh E2E validation (finops-mesh ATLAS + multistep) to follow on this PR.

Test plan

  • finops-mesh ATLAS run with aliased objectives succeeds (no schema error)
  • run completes promptly; FLAG captured

- generate_atlas_attack: normalize objective aliases (id<-name/objective_id,
  goal<-description/objective/prompt) and give a per-object error naming the
  missing field, instead of rejecting near-miss schemas with a terse message.
- _auto_execute_workflow: lower the default subprocess timeout 3600s -> 1200s
  as a backstop so a stalled run can't block the tool for a full hour.
- bump version 1.20.1 -> 1.20.2 for manifest sync.
@rdheekonda
rdheekonda merged commit a970acb into main Oct 27, 2026
5 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