Skip to content

fix(web-security): add continuous STATUS emission and Unexplored fiel…#54

Merged
GangGreenTemperTatum merged 1 commit into
mainfrom
ads/eng-7147-fix-web-capability-agent-prompt-directives-drift
Jun 13, 2026
Merged

fix(web-security): add continuous STATUS emission and Unexplored fiel…#54
GangGreenTemperTatum merged 1 commit into
mainfrom
ads/eng-7147-fix-web-capability-agent-prompt-directives-drift

Conversation

@GangGreenTemperTatum

Copy link
Copy Markdown
Contributor

…d to prevent exploitation drift

Three targeted changes to the web-security agent prompt:

  1. Act step now mandates emitting a STATUS line immediately after every action, not only at named milestones (recon, surface completion, conclusion). Continuous emission makes state drift visible before the next action is chosen, preventing silent re-entry into surface scanning.
  2. STATUS format gains a mandatory Unexplored field listing attack surfaces observed but not yet deeply tested. A non-empty Unexplored field is an explicit signal that the agent has not reached genuine exhaustion and must return to Decide before concluding.
  3. Attack Methodology clarifies that the vulnerability checklist is the minimum bar, not the ceiling. Once standard classes are covered the primary question becomes gadget composition — the most impactful findings are novel chains, not textbook one-shot vulnerabilities.

…d to prevent exploitation drift

Three targeted changes to the web-security agent prompt:

1. Act step now mandates emitting a STATUS line immediately after every
   action, not only at named milestones (recon, surface completion,
   conclusion). Continuous emission makes state drift visible before the
   next action is chosen, preventing silent re-entry into surface scanning.

2. STATUS format gains a mandatory `Unexplored` field listing attack
   surfaces observed but not yet deeply tested. A non-empty Unexplored
   field is an explicit signal that the agent has not reached genuine
   exhaustion and must return to Decide before concluding.

3. Attack Methodology clarifies that the vulnerability checklist is the
   minimum bar, not the ceiling. Once standard classes are covered the
   primary question becomes gadget composition — the most impactful
   findings are novel chains, not textbook one-shot vulnerabilities.
@GangGreenTemperTatum GangGreenTemperTatum merged commit 4ca2f0d into main Jun 13, 2026
4 of 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