Hey team, per Vasco's suggestion, I'm dropping a parallel development relevant to the SDK's longer-term security posture in this channel. I've joined the AARM (Autonomous Action Runtime Management) working group chaired by Herman Errico at Vanta. AARM is the agent governance standard Vanta donated to CSA on April 30 (https://aarm.dev). The SDK's existing security analyzer (baseline in PR #2709, merged 12 May) maps to AARM Core: •

PatternSecurityAnalyzer

/

PolicyRailSecurityAnalyzer

/

EnsembleSecurityAnalyzer

→ Policy Engine •

ConfirmRisky

+

confirmation_policy.py

→ Approval Service •

_extract_*

and

_normalize

in

utils.py

→ partial Action Mediation Layer •

SecurityRisk.UNKNOWN

semantically maps to AARM's DEFER verdict • Context Accumulator and Receipt Generator are honest gaps The SDK's ensemble composition surface and the UNKNOWN-on-ERROR fusion policy under discussion in #2721 aren't in the AARM spec at all. AARM doesn't yet prescribe multi-evaluator composition rules. I'd like to submit that pattern back to AARM as a vendor-neutral contribution, since the SDK is ahead of the spec there. That opens a larger question: how does OpenHands as a project want to engage with AARM, if at all? I'm not qualified to walk through that door as a humble contributor. I'm looking for feedback here - I am happy to keep my engagement with AARM separate from OpenHands. Though, the opportunity was worth mentioning.

Very interesting spec, thank you! About the project's potential engagement, cc: @wide-doctor-37830 @limited-student-10747

Hi @refined-secretary-75241, this is very exciting! I think at least in principle this is something that the maintainers are pretty interested in, but it's a question of effort scheduling. What I will say is that you can definitely speak up as an OpenHands contributor, since you know our efforts in this direction well. I'll also ask around if any of the maintainers would be able to join as well.

@limited-student-10747 thank you - I'll keep y'all posted as I participate in the working group. I understand the effort scheduling constraint, so no worries.