From Fragmented Signals to Collective Intelligence: BureauID and Silence Laboratories to Build a New Foundation for Private and Collaborative Fraud Prevention

BureauID and Silence Laboratories partner to enable collaborative fraud prevention without sharing sensitive customer data.

Portrait of Kush Kanwar

Written by

Kush Kanwar

Partnerships

Oct 26, 2026

Fraud prevention is no longer just about how much an individual institution knows. It is about how effectively institutions can connect intelligence across an ecosystem, without compromising the sensitive data behind it. BureauID and Silence Laboratories are working together to make that possible.

BureauID brings its AI-powered identity and risk decisioning platform, connecting identity, device, behavioural and transaction signals to help businesses verify digital identities, detect fraud and monitor risk in real time. Its platform serves more than 200 banks, fintechs and online businesses, enabling risk teams to make better decisions through AI and no-code workflows.

Silence Laboratories provides the privacy-preserving computation infrastructure that enables multiple organisations to collaborate on sensitive datasets without exposing raw information. Through its cryptographic encrypted computing technology, built on multiparty computation (MPC) and zero-knowledge proofs, organisations can jointly run analytics, detect patterns and exchange actionable intelligence while maintaining control over their underlying data.

Together, the two aim to unlock a richer layer of collective fraud intelligence: connecting previously fragmented signals across financial institutions and industries, enabling cross-network fraud detection, stronger entity resolution and more informed risk decisions, without requiring raw customer data to be shared between participants.

Consider a mule account that appears entirely ordinary to the bank holding it. A telco knows its phone number moved to a new SIM yesterday. A fintech has seen the same device open five accounts this month, while a lender has seen the same address, lightly disguised, on three applications.

Each fragment looks harmless. Together, they point to a fraud ring.

The challenge is that fraudsters operate across institutional boundaries, while the intelligence used to stop them remains fragmented. The result is missed patterns, delayed investigations, unnecessary false positives and fraud networks that can exploit the gaps between organisations.

The global picture has long been grim. The UN Office on Drugs and Crime has estimated that criminals laundered around $1.6 trillion, or 2.7% of world GDP, while less than 1% of illicit financial flows were being seized and frozen. And each incident costs more than the money taken: investigations, teams pulled into crisis response, and customers who lose trust and leave.

Figure 1: Four institutions, four fragments: only joined up do they point to a fraud ring.

Why fraud detection is a collective effort

Collective intelligence means reading signals from every touchpoint and every industry a customer deals with as one picture of risk. Fraudsters already work across institutions, probing one set of defences and moving on to the next. Defenders need the same reach, combining onboarding, device, login and payment signals from banks, telcos, fintechs, lenders and online merchants. Joined up, isolated alerts become a holistic view of the customer, with fewer false positives and less friction for genuine users.

It helps to look along two lines. Along the customer journey, five kinds of signal matter: identity (who is this person?), device (what are they using?), behaviour (how are they acting?), network (who else is connected?) and transaction (where is the money going?). One signal is a hint; together, they reveal a pattern.

Figure 2: Five kinds of signal along the customer journey. Scams cross calls, messages, devices and accounts, so no single industry sees them all.

Across industries, each sector holds a different part of the same person. Telcos see SIM age, recent swaps and recycled numbers. Banks see account age, money flows and new payees. Wallets and fintechs see devices shared across accounts. Marketplaces see shared delivery addresses and refund abuse. Lenders see applications made elsewhere and loan stacking, and insurers, gaming and gig platforms see repeat claims and multiple accounts. The strongest signal for a bank may sit in another industry altogether.

Figure 3: Each industry sees a different part of the same person. A recent SIM swap or a shared delivery address can change a bank’s assessment.

The patterns that matter most are invisible from inside one organisation. In circular transactions, money returns to where it started after passing through several institutions. Mule accounts collect from many unrelated senders and quickly pass the money on. In layering, value splits, hops between institutions and merges again. Launderers also split cash deposits across banks to stay under reporting thresholds, and synthetic identities apply for credit at many lenders at once. Seen together, these patterns stand out; seen apart, each looks routine.

Figure 4: Circular transactions, mule accounts and layering only become visible when institutions’ views are joined.

Regulators are pushing towards collective fraud intelligence

This need for ecosystem-wide intelligence is increasingly reflected in regulatory and industry initiatives around the world. Authorities are recognising that institution-level monitoring alone is insufficient for combating sophisticated fraud and financial crime, and are encouraging frameworks that enable information sharing and coordinated detection.

In Singapore, the Monetary Authority of Singapore launched COSMIC to facilitate information sharing among participating financial institutions when specified financial-crime risk thresholds are met. In India, the Reserve Bank of India has advanced the Digital Payments Intelligence Platform (DPIP) to strengthen payment fraud detection through shared intelligence, alongside the Reserve Bank Innovation Hub’s MuleHunter.ai initiative for identifying mule accounts. In the United States, Section 314(b) of the USA PATRIOT Act provides a voluntary legal framework for eligible financial institutions to exchange information related to suspected money laundering and terrorist financing, including qualifying fraud-related activity.

While these initiatives differ in scope and implementation, the direction is consistent: fraud and financial crime prevention increasingly require coordinated action across institutions, not just stronger defences within them.

However, enabling information sharing is only one part of the challenge. The next question is how institutions can collaborate effectively without creating new risks around customer privacy, commercial confidentiality, data sovereignty and regulatory compliance.

Why sharing stalls

The pieces stay apart because sharing raw data is risky for everyone involved. The data that catches fraud is sensitive, and four kinds of barrier keep it at home:

  • Legal. Privacy rules, sector regulations and cross-border restrictions limit how data can move, and each new data-sharing agreement can take months of legal work.

  • Commercial. Customer data is a competitive asset, so members need a reason to contribute and a fair share of the value.

  • Technical. Identities must be matched across systems without exposing customer records.

  • Trust and liability. Someone has to answer for pooled data, the quality of investigations and any misuse.

The usual options each fail. Share raw data, and copies leave your control, each one a new exposure. Keep it siloed, and fraud across institutions goes unseen. Hand it to a central party, and you create a single point of failure and of trust.

The common patches don’t fix this. Hashing fails because there are only a few billion possible mobile numbers; one graphics card can hash them all and match them back in minutes. Anonymisation fails the other way: strip out the identifiers and you strip out the links that expose a ring. Useful collaboration needs strong privacy without losing the signal.

Figure 5: The usual workarounds trade usefulness for privacy. Privacy-enhancing technologies don’t have to.

Agentic AI makes this problem even more critical

As financial services become increasingly automated, collaboration will also become more agentic.

AI agents will increasingly participate in customer onboarding, payment approvals, fraud investigations and risk monitoring. Rather than relying solely on manual information requests, an institution’s agent could query agents or services across several organisations, combining intelligence and acting on the results in real time.

Imagine a bank’s onboarding agent checking with a telco whether a customer’s number was recently SIM-swapped, while simultaneously checking with a fintech whether the associated device has links to known mule accounts. A payment agent could similarly assess the receiving account’s cross-network risk before releasing a transaction.

Such workflows could make fraud prevention faster, more contextual and increasingly proactive.

But agentic collaboration also compounds the privacy problem.

As agents exchange information across organisational boundaries at machine speed, every additional interaction can create another potential point of exposure. Agents may inadvertently reveal sensitive information, access more context than necessary or be manipulated into disclosing data beyond their authorised purpose.

In the MAGPIE benchmark, leading models leaked up to 50.7% (Gemini 2.5 Pro) and 35.1% (GPT-5) of the sensitive information in multi-agent tasks, despite explicit instructions not to.

In PrivacyLens, GPT-4 and Llama-3-70B leaked sensitive information in 25.68% and 38.69% of cases. The studies measure different settings, but the lesson is the same: a prompt cannot enforce a privacy boundary.

Simply instructing an agent not to share sensitive information is not a sufficient privacy guarantee.

In an increasingly agentic ecosystem, privacy cannot be an afterthought or a policy enforced only through prompts. It must be built into the computation and collaboration infrastructure itself.

Private agent collaboration therefore needs three things, built into the infrastructure rather than the prompt:

  • Answers without raw records. Agents query across the network and receive only the permitted answer, so raw customer data never enters the agent exchange.

  • Policies enforced at runtime. The system checks access and computation rules on every call, and enforces only the permitted ones.

  • Authorised and auditable. Every question needs authorisation and leaves a record, so members can verify what ran and what it revealed.

Pool the insight, not the data

Privacy Enhancing Technologies (PETs) let institutions pool their insight without pooling their data. A bank, a telco and a fintech can jointly ask, “has this device, number or account appeared in a fraud case anywhere in the network?” Each keeps its records in-house, and only the answer comes out: a match, a score or a flag.

Figure 6: A member asks an authorised question; the computation runs across every member’s encrypted data, and only the answer comes back.

A yes or no is only the start. Queries can combine several identifiers, match names and addresses even when they have been lightly altered, apply custom conditions, produce shared risk scores, run network analysis and screen many customers in parallel. Every query can be logged and justified, and supervisors can observe the network without seeing customer data.

In practice, it works in four steps:

  1. Connect. Each member’s data stays encrypted on its own systems, and members agree which fraud questions the network may answer.

  2. Compute. Approved detection queries run across all members’ data at once.

  3. Detect. The joined-up view exposes mule networks, layering and circular flows.

  4. Act. Each member receives alerts about its own accounts only, and decides whether to investigate, report or intervene.

Here is what each member puts in and gets back for the three patterns above:

Table 1: Illustrative checks and outputs. What each member sees follows the permissions agreed by the participants.

Nobody, including whoever runs the system, sees anyone else’s data. Privacy is enforced by maths, not just contracts, and every new member makes each check sharper. The deep dive below explains how.

Figure 7: Agents that swap records multiply exposure; agents that ask through a joint check get answers and leave the data where it is.

Silence Laboratories and BureauID

This is the foundation BureauID and Silence Laboratories are working together to establish.

BureauID brings the identity intelligence, real-time risk signals and decisioning capabilities needed to identify suspicious behaviour and act on risk. Silence Laboratories brings the cryptographic computation layer that allows those capabilities to extend across organisational boundaries while protecting the underlying data.

Together, the collaboration aims to enable financial institutions and other ecosystem participants to enrich existing fraud detection with intelligence from previously disconnected sources.

The potential applications extend from private fraud-list screening and cross-network entity resolution to detecting mule networks, identifying coordinated fraud patterns and enabling collaborative risk assessment.

And as fraud prevention becomes more agentic, the same infrastructure can support secure agent-to-agent intelligence exchange, with privacy, authorisation, policy enforcement and auditability built into the computation rather than dependent on agents’ instructions.

A practical starting point is a focused consortium of participating institutions, an agreed set of fraud detection queries and a framework for returning relevant alerts to each participant without exposing anyone else’s customer data.

Fraud networks already operate across institutional boundaries. The intelligence used to stop them needs to do the same without the underlying sensitive data following it.

Technical deep dive: MPC and federated learning

Several cryptographic tools make this possible.

Multi-party computation (MPC) lets several parties compute a joint result while each input stays private. Inputs are split into random-looking fragments, or secret shares, held by independent nodes, often run by the participants themselves. No node learns anything on its own, yet together they compute on the shares and reveal only the final output.

For fraud teams, MPC powers private matching (is this number on any member’s fraud list?), cross-bank totals that expose deposits split to dodge reporting thresholds, and network analysis that traces mule chains among other things.

Here is the idea in brief. Three banks each saw the same customer deposit just under $10,000 in cash this week: $9,000, $8,500 and $9,200. Each bank splits its figure into three random-looking shares, one per node. Each node adds up the shares it holds, which mean nothing on their own. Combining the three partial sums reveals only the total, $26,700, a classic structuring red flag.

Figure 8: Each node holds one column of meaningless shares; adding up the column totals reveals only the sum.

Federated learning (FL) trains a shared model without centralising data. Each institution trains on its own records and sends only model updates to be combined. Those updates can still leak details of the underlying data, so FL is paired with secure aggregation, an MPC protocol that reveals only the sum of everyone’s updates.

Figure 9: Each member trains locally; secure aggregation combines the masked updates so only their sum is ever revealed.

The two complement each other. FL learns what fraud looks like across the network over time; MPC applies that knowledge to live, joined-up signals at the moment of decision.

No headings found on page

SHARE

Continue reading