Skip to content

fix(ai-red-teaming): scope get_session_context to AIRT + make it pull-only (CAP-1169) - #109

Merged
rdheekonda merged 1 commit into
mainfrom
fix/cap-1169-scope-session-context
Jul 30, 2026
Merged

fix(ai-red-teaming): scope get_session_context to AIRT + make it pull-only (CAP-1169)#109
rdheekonda merged 1 commit into
mainfrom
fix/cap-1169-scope-session-context

Conversation

@rdheekonda

Copy link
Copy Markdown
Contributor

Closes CAP-1169.

Problem

get_session_context's docstring said "use this at the start of follow-up commands… so you can build on previous attacks" — a push instruction that primed agents to open with it. When the AIRT capability is co-attached with another (e.g. web-security), the agent reaches for this AIRT tool first even on non-red-teaming primitives → occasional spins + odd demos. The goal is to stop that without temporarily disabling AIRT.

Change (priming fix — pull-only + AIRT-scoped by description)

  • Reword the get_session_context docstring: pull-only, red-teaming-only, resume aid — call ONLY to continue a prior attack (try another attack / add transforms); do NOT call to start a task; not relevant to non-red-teaming work.
  • Soften the empty return: "No prior red-teaming session. Nothing to resume — proceed with the task directly." (was "Run an attack first…", which invited a spin).
  • Scope the agent-prompt tool description to match.
  • Patch bump 1.10.0 → 1.10.1.

No behavior change when genuinely resuming an AIRT attack.

Scope note (soft vs hard)

This is a description/priming fix — the tool is still present in the toolset, but its wording steers the agent away from it for non-AIRT tasks. If we want it hard-hidden when a non-AIRT capability is selected, that's a follow-up at the capability-loader level (conditional tool exposure by active capability), not something a docstring controls.

…-only (CAP-1169)

get_session_context's docstring said "use this at the start of follow-up
commands", which primed agents to open with it — including on non-red-teaming
tasks when the AIRT capability is co-attached (agents spin on non-AIRT
primitives / odd in demos). Reword to pull-only, red-teaming-only, resume-aid
wording (do NOT call to start; not relevant to non-red-teaming work), soften
the empty-return to "proceed with the task directly", and scope the agent-prompt
description accordingly. No behavior change when genuinely resuming an attack.
@rdheekonda
rdheekonda merged commit fd25699 into main Jul 30, 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