3
independent intelligence pipelines
0
raw records ever leave your perimeter
0
raw data shared between members, ever
60s
cadence on signed rule distribution
The asymmetry problem

Attackers are already federated. Defenders rediscover the same attack class on their own.

The 2025 attack surface confirmed it in detail. The same prompt-injection patterns appearing in CrowdStrike's 2026 Global Threat Report were observed against multiple unrelated banks within days of each other. The malicious postmark-mcp campaign was identified independently by individual victims, not by the security industry. By the time public disclosure occurred, downstream agents had been routing email through the attacker's server for weeks.

Problem 01

FS-ISAC's IOC formats predate AI agents

FS-ISAC remains the gold standard for inter-bank security collaboration, but its IOC formats and threat advisories were designed for IPs, hashes, domains, and malware families. They do not express the artifacts of an AI agent attack: prompt structure, tool-call sequences, MCP endpoint hashes, behavioral baselines, or origin-tagged context fragments.

Problem 02

Privacy and competitive concerns block raw sharing

Prompts contain non-public personal information. Tool-call parameters may contain account numbers. Sharing raw artifacts violates GLBA Reg P and is unacceptable on competitive grounds. The result: every institution rediscovers the same attack class on its own. The Federation Network is the architecture that breaks that constraint.

Three independent intelligence pipelines

Federation rules are not generated solely from member-bank signal.

The Federation Service ingests three independent pipelines, applies the same privacy-preserving distribution architecture to all of them, and tags every published rule with its provenance, so a member bank's threat team can distinguish research-derived rules from peer-observed rules at a glance.

BladeRun Threat Research

A dedicated adversarial AI research function, a small team of senior human researchers directing a fleet of internal red-team agents that probe BladeRun's own detection stack, mine adversarial-ML literature, and generate technique variants at scale.

Reference architecture is OverWatch, Mandiant Threat Intelligence, and Recorded Future Insikt Group, applied to AI agent attack tradecraft. Every published rule carries a named human researcher's signoff in its provenance metadata.

Day-one valueHuman-signedMITRE ATLAS-mapped

BladeRun Honeypot Network

Instrumented decoy MCP endpoints, decoy public-facing agents, and seeded indirect-injection content. Any attacker probing them produces a Federation rule with the cleanest privacy posture of any signal source, synthetic data only, no real customer data in the source feed at any point.

No GLBA, NPI, PCI, or GDPR handling obligation attaches to honeypot-derived rules. Rules are tagged with provenance for per-source review policy.

Day-one valueSynthetic-data onlyCleanest privacy posture

Cross-Member Correlation

Anonymized signal hashes contributed by member banks under a k-anonymity floor, with blind-index member identity, salted hashing, and differential privacy on baseline gradients. Exact parameters are specified in the Federation white paper.

A Federation rule is generated only when k or more independent members observe the same signal. Member identity is recoverable only by the member, via a key held in the bank's HSM. BladeRun staff cannot link signal to bank.

Compounds with membersk-anonymityHSM-held identity key

A bank that joins on day one receives the existing research-derived and honeypot-derived rule corpus immediately, with no member-correlation dependency required to deliver value. As the network grows, cross-member correlation adds a third layer that compounds the others.

What each bank contributes

Four classes of one-way signal. Each one cryptographically pre-image resistant.

A contributing bank cannot reconstruct the original event from the published artifact. The Federation Service cannot reconstruct it either. Every signal class is documented with its hash inputs, privacy property, and contribution semantics.

Signal classWhat the hash encodesPrivacy property
Injection-pattern signal hashA salted hash of structural features extracted by the injection classifier, never the prompt content itself. The original prompt text never leaves the bank.Pre-image resistant · collision-resistant
MCP endpoint hashA salted hash of endpoint fingerprint attributes for any MCP endpoint that triggered a registry violation. Allows cross-institution correlation of typosquat campaigns without revealing the bank's approved endpoint list.URL not recoverable from hash
Tool-call anomaly fingerprintBucketed structural anomaly: agent type, tool name, parameter format anomaly class, anomaly score band. No customer identifiers, no parameter values, no business context.k-anonymity ≥ 50 enforced before publishing
Federated baseline gradientPer-agent-type behavioral baseline updates encoded as model-weight deltas under differential privacy. No raw call sequences are transmitted; only differentially-private weight updates participate in the global baseline.(ε, δ), differential privacy
What each bank receives

Five artifact classes. Pulled by the Federation Client every 60 seconds.

Every artifact arrives signed and timestamped. Every rule is tagged with provenance, Research, Honeypot, or Member-Correlation, so the bank's threat team can apply different review and adoption policies to each source class.

01 / Detection rules

Detection rules

Member-correlation rules are issued when the same injection-pattern hash appears across enough independent members within a bounded time window. Research and honeypot rules are issued when BladeRun's Threat Research team validates a signature against published precision and recall thresholds.

02 / Prevalence telemetry

Industry prevalence

Per attack class, the count of unique members observing that class in the trailing window, never the identities of those members. A CISO can answer "is this attack hitting our peers?" without any peer being named.

03 / Policy updates

Recommended policy updates

When a new attack class warrants a default-policy change (a new MCP-endpoint validation rule, for instance), the recommendation is distributed for human review at each member bank. Banks adopt or reject. No automatic policy change is forced.

04 / Model updates

Federated model updates

The injection classifier and the behavioral baseline models receive periodic global-aggregate updates. Each member's local model improves with cross-institution data without any raw cross-institution data ever existing in one place.

05 / ATT&CK mapping

MITRE ATT&CK / ATLAS mapping

Every Federation rule is tagged with its MITRE ATT&CK and ATLAS technique IDs, enabling direct integration with the bank's existing detection engineering tooling.

06 / Quarterly briefings

Quarterly Threat Landscape

Live briefings to member-bank threat teams. Format follows the quarterly threat reports your team knows from CrowdStrike, Mandiant, and Recorded Future, scoped to AI agent attack tradecraft.

The privacy contract

What never leaves the bank's perimeter.

This is the section your privacy, legal, and information-security teams will read first. Every item below is engineering-enforced, not policy-promised. The on-prem Federation Client is open for review.

ArtifactStatus
Raw prompt text (input to LLM)STAYS, never serialized to Federation Client
LLM response textSTAYS, never serialized
Tool call parameter values (account numbers, amounts)STAYS, only structural anomaly class is published
Customer identifiers, names, NPI, PCI dataSTAYS, Presidio redaction occurs before any signal extraction
Bank-specific business logic, agent prompts, system instructionsSTAYS, never exposed in any form
Approved MCP endpoint list, internal API surface, partner integrationsSTAYS, only violation hashes (denied connections) are shared
Identity of which other banks are membersSTAYS, membership is double-blind, even from BladeRun staff
Two topologies, one coverage envelope

Bank-built agents calling out. External agents calling in.

Both directions are real production patterns at every Tier 1 bank we've spoken with. Both are inside BladeRun's coverage envelope. They look different in the call diagram and they look different in what we can claim.

Topology A · Outbound

The bank built the agent

Prompts, tool definitions, orchestration code all run inside the bank's perimeter; the agent calls outbound to an external LLM. We see everything: prompt, tool calls, sub-agent spawns, response.

Bank Agent → Sentinel SDK → Gateway →
External LLM (OpenAI / Claude / Bedrock / …)
Coverage: full call shape, prompt content, tool calls, response content
Topology B · Inbound

An external agent calls in

Claude, ChatGPT, Operator, Gemini, or a third-party agent calls inbound into the bank's MCP servers or APIs. We see what we see, we do not see what we do not see, and our cross-institution correlation matters most.

External Agent (Claude / Operator / Gemini / …) →
Gateway → Bank MCP / API surface
Coverage: call shape, signed agent identity, mandate scope, behavioral baseline

The InAuth precedent is direct: device fingerprinting won the mobile-banking decade by scoring callers at the bank's boundary without needing visibility into the device. BladeRun applies the same pattern at the agent boundary.

Three claims, defensible under scrutiny

What a privacy or compliance review will challenge, and our answer.

Federation Network · Audit-grade summary

  1. Federation rules are sourced from three independent intelligence pipelines, BladeRun Threat Research, the BladeRun Honeypot Network, and Cross-Member Correlation, and member banks see provenance on every rule.
  2. Member banks contribute only one-way attack-signal hashes and pattern metadata, never prompts, never customer data, never raw logs, and receive detection rules, prevalence telemetry, and federated model updates that strengthen every individual deployment.
  3. An LLM provider, an EDR vendor, an identity vendor, or a research-only AI security vendor cannot replicate this network. None of them have runtime presence in the AI agent call path across multiple banks simultaneously combined with a research and honeypot intelligence function.
Founding members

Three to five Tier 1 banks for 2026.

Founding members shape the detection model that every future member inherits. Direct access to founders. Design partner pricing. Founding seat in the Federation governance body, including the privacy contract, the rule-promotion thresholds, and the cadence of human-reviewed policy recommendations.

Step 01

Signed NDA + technical session

One-hour meeting covering the full architecture, the privacy contract, and the Federation Client's open-source structure. Your privacy and infosec teams in the room.

Step 02

Privacy + compliance review

We provide the documentation pack: data flow diagrams, hash specifications, k-anonymity proofs, DP parameters, and the Federation Client source for your team's audit.

Step 03

Local engine first, Federation later

Local BladeRun deployment ships with the Threat Research and Honeypot rule corpus from day one. Federation contribution is opt-in and reversible.

Step 04

Federation Client deployment

Pull-based, signed-update channel. 60-second cadence. Your local rule engine begins receiving cross-pipeline rules. Contribution begins when your privacy team signs off.

Works with

The Federation rides on the BladeRun runtime stack.

Common questions

What privacy, legal, and security teams ask first.

Three controls: (1) Presidio redaction occurs before any signal extraction. NPI never reaches the hashing pipeline. (2) Only the four signal classes documented in the white paper are emitted, and each is cryptographically pre-image resistant. (3) The Federation Client is open for your team's review; you can audit exactly what is serialized.
No. Membership is double-blind by design. Other members do not know which banks are participating. BladeRun staff do not have a list that links signal contributions to specific members. The linkage is held in your HSM, recoverable only by you. Industry-prevalence telemetry counts unique members but never names them.
A k-anonymity floor governs when a member-correlation rule is issued, with a substantially higher floor before tool-call anomaly fingerprints are published. Founding-member note: at small network sizes, peer-attribution resistance is limited; the property strengthens as the network grows past 20+ members. The white paper specifies the threshold parameters in detail.
Complementary. FS-ISAC remains the gold standard for traditional threat intelligence sharing, IPs, hashes, domains, malware. The Federation Network covers what FS-ISAC's IOC formats were not designed for: AI agent attack tradecraft. We expect Federation rules to feed into bank threat-intel platforms alongside FS-ISAC advisories and CrowdStrike intelligence.
Yes. Federation participation is opt-in per signal class. A bank can subscribe to the Threat Research and Honeypot rule corpus while opting out of cross-member correlation contribution. The economics may evolve over time, but the architecture supports asymmetric participation.
Three structural reasons. (1) Each provider sees only its own slice of the bank's traffic. They cannot aggregate cross-provider AI-runtime signal. (2) They are direct competitors with each other; OpenAI will not inspect Anthropic calls. (3) None of them has runtime presence in the AI agent call path across multiple banks simultaneously, combined with a dedicated research and honeypot intelligence function. Federation requires sitting where they don't sit.
Founding members for 2026

Three to five Tier 1 banks. Direct access to founders.

Your threat model shapes the detection corpus. Your privacy review shapes the Federation contract. The cohort closes at five.