PRA SS2/21
PRA Supervisory Statement SS2/21 — Outsourcing and Third-Party Risk Management
PRA third-party risk-management expectations. LLM providers and agent runtimes are third parties under SS2/21; concentration risk is explicit.
Who does it apply to?
- UK PRA-regulated firms with material outsourcing or third-party arrangements.
- Any PRA-regulated firm using a public LLM provider (OpenAI, Anthropic, Google Vertex AI) — these are third-party providers under SS2/21.
- Firms running agents on third-party agent runtimes — Vertex AI Agent Engine is a third-party service even though it's GCP-native.
Two-minute explainer
PRA Supervisory Statement SS2/21 became effective on 31 March 2022. It sets PRA expectations for how UK PRA-regulated firms manage outsourcing and third-party arrangements. For agentic AI builders, SS2/21 matters because the LLM provider, the cloud provider, and any third-party API the agent calls are third parties under the statement.
Material outsourcing is the threshold concept. SS2/21 distinguishes material arrangements — those that, if disrupted, would have a material adverse impact on the firm’s safety, soundness, or operational resilience — from non-material. Most regulated firms running agentic AI workflows in production will classify their LLM provider as a material arrangement: the agent doesn’t work without the model.
Once classified as material, several obligations attach:
- Pre-contract due diligence. Documented assessment of the provider’s capability, control environment, and operational resilience.
- Concentration risk monitoring. Over-reliance on a single provider is a regulator concern. SS2/21 doesn’t prescribe a single threshold; it expects the firm to define one. The Regulus Model Registry’s concentration metric is the runtime view.
- Exit planning. A documented exit strategy that’s regularly rehearsed. Regulus’s kill-switch plugin doubles as the exit-drill mechanism — engaging the switch on the primary LLM provider exercises the runbook.
- Sub-outsourcing transparency. If your LLM provider sub-outsources any material function (e.g. compute capacity), that needs to be identified. Most LLM providers’ sub-outsourcing chain is opaque; Regulus tracks what’s disclosed but can’t conjure transparency from nothing.
- Ongoing monitoring. Continuous monitoring of the third party’s performance against the SLA, incident handling, and audit rights. The audit chain captures the firm’s side of the relationship.
The Regulus pra-ss2-21 profile is intentionally narrower than the
SS1/23 profile — most of SS2/21 is governance-side documentation, not
runtime evidence. What Regulus delivers: the material-arrangement
register, the concentration metric, the exit-drill audit trail, and
the ongoing-monitoring substrate. The actual contract terms, the SLA
text, the audit-right enforcement letters — those live in your
procurement and vendor-management systems.
The profile composes naturally with pra-ss1-23 (model risk is a
specific case of third-party risk where the third party is an LLM),
with dora (DORA’s ICT third-party register has overlapping field
expectations), and with fca-sysc (FCA SUP 16 reporting on outsourcing
runs in parallel for dual-regulated firms).
What it actually requires of an engineer
- Material third-party arrangements need explicit identification. The Model Registry tracks LLM provider as a third-party with the SS2/21 metadata.
- Concentration risk is a real obligation. Over-reliance on a single LLM provider — measured as % of inference, % of revenue dependency, % of critical operations — triggers concentration-risk monitoring.
- Exit planning must be evidenced. What's the alternative model provider? Can the agent fail over? How fast? Document the runbook; the audit chain captures exit-drill events.
- Reverse-stress-test scenarios apply to LLM providers. If your primary LLM provider went offline / changed pricing / changed terms, what's the operational impact? Regulus emits the data points; the analysis is governance.
What Regulus does for you
| Regulus control | Delivers |
|---|---|
RegulusModelRiskPlugin | Third-party register entries — every LLM provider is a registered third party with material-arrangement classification, concentration metric, and exit-runbook reference. |
RegulusGovernanceEvidencePlugin | Quarterly third-party register export with SS2/21 fields — material-arrangement classification, concentration metric % per provider, exit runbook reference, sub-outsourcing chain. |
RegulusKillSwitchPlugin | Exit-drill capability — engaging the kill switch on a primary LLM provider exercises the failover runbook. Audit chain captures the drill event with the failover region, time-to-recovery, and the alternative provider. |
RegulusIdentityExpiryGuard | Third-party access credentials must rotate within the SS2/21 control framework. Expired LLM-provider credentials fail closed. |
Saves you ~8 engineer-weeks
Estimate based on the following honest assumptions:
- Material-arrangement classification logic (2 weeks).
- Concentration metric calculation + monitoring (2 weeks).
- Exit-drill automation + audit (2 weeks).
- Quarterly register export in SS2/21 format (2 weeks).
What an auditor will ask
The questions you'll see in a real walkthrough — and where to point them.
-
Show me your material third-party register.
regulus third-party register export --format ss2-21produces the register. The Model Registry is the source; the export applies the SS2/21 field overlay. -
What's your concentration risk on LLM providers?
The concentration metric in the Model Registry shows % of inference per provider. Configurable thresholds (typically 40% / 60% / 80%) trigger CONCENTRATION_THRESHOLD_BREACH events with the date and exposure level.
-
Walk me through your exit plan for the primary LLM provider.
The Model Registry has an
exit-runbookreference per provider. The runbook is your governance artefact; the audit chain shows the most recent exit drill — date, failover region, time-to-recovery, alternative provider used.
What this doesn't cover
- Non-AI outsourcing — Regulus profiles AI agent third-party dependencies; other outsourcing (back-office processing, market-data vendors) is your operational-risk function's scope.
- Cloud-provider outsourcing as an organisational arrangement — Regulus tracks the AI side; the broader GCP outsourcing arrangement is at the firm level.
- Cross-border data flows — covered by GDPR / UK GDPR residency, not SS2/21.
Citations
Activate this profile in your agent
regulus init my-agent --profiles=pra-ss2-21