Wall Check: tests for agents that piece together inside information across a bank's information barrier
Red-team tests that check whether internal AI agents assemble inside information from scattered documents. Broker-dealer compliance teams pay.
Early · Agent Mosaic measures assembled leaks
Written for: Small startup team
Next step
Call the head of the control room at a mid-size investment bank such as Stifel or Piper Sandler. Ask whether their document-agent or Copilot rollout was ever tested for agents piecing together inside information across the wall, and who would sign off on a test.
Why now
Banks are putting agents over deal files: V7 Go already screens investment deals across thousands of documents, and the Anthropic–Hg consultancy exists to speed up rollouts like these. Section 15(g) of the Exchange Act already requires broker-dealers to have written policies that stop inside information being misused. An agent that sits on the public side but reads shared drives can break that policy without opening anything restricted. Control rooms test people against this rule today, not agents. Nobody sells a mosaic test for it, so the space is open.
What became possible
Agent Mosaic measures the 'mosaic effect': an agent reads documents it is allowed to see, then discloses a sensitive conclusion built from the pieces, even when the question is benign. The authors ran 30 synthetic cases, producing 10,800 answers from three agents in simulation and 1,440 from live agents in two open-source harnesses.
Two ways in
- Wall Check
A test harness that seeds synthetic deal documents (pieces of a merger, a financing, an earnings miss) across the folders a public-side agent can read. It then asks the agent ordinary research questions and scores whether its answers assemble the restricted conclusion.
- Customer
- Control room and compliance testing at mid-size investment banks and broker-dealers
- Model
- Fixed-fee assessment per agent deployment, then an annual retest subscription
- Mosaic Monitor
A runtime check on agent answers that flags any answer naming or implying an issuer on the firm's watch or restricted list. Flagged answers are held for control-room review.
- Customer
- Same control rooms, after the first assessment
- Model
- Per-agent monthly fee
Competition
No product was found that tests whether a bank's internal AI agents piece together inside information across the information barrier. Incumbents such as Intapp enforce access permissions, which does not catch an agent combining documents it is allowed to read.
Full competition notes (9 competitors)
No product was found that tests whether a bank's internal AI agents piece together inside information across the information barrier. Incumbents such as Intapp enforce access permissions, which does not catch an agent combining documents it is allowed to read. AI security vendors such as Adversa and DeepInspect do not ship finance scenarios built around the restricted list. Academic work, mainly the Agent Mosaic paper published days ago, proves the problem is real and measurable but uses non-finance scenarios. A new entrant can win by owning the finance-specific harness: realistic synthetic merger, financing and earnings documents, scoring tied to restricted-list and wall-crossing logic, and audit evidence a control room can show examiners under the broker-dealer rules on misuse of inside information. Skadden's July 2026 alert supplies the regulatory hook. The domain detail is the moat, because the basic method is now public. Mosaic Monitor, the runtime check, is weaker on its own. Checking answers against the restricted list is an easy feature for Intapp, DeepInspect, SoftwareOne or Copilot guardrail vendors to add, so it is best sold to customers who already use Wall Check. The main risk is Intapp adding a testing module, since it already sits inside investment banks' wall infrastructure.
- Intapp Walls for AI / Celeste AI Wall
An ethical-wall product that applies the same permissions to AI tools as to staff. It is sold mainly to law firms, integrates with Harvey and Claude, and Intapp also has an investment-banking line. It stops agents reading documents they shouldn't, but it does not test whether an agent pieces together a restricted conclusion from documents it may read. Most likely to add this capability.
- Intapp Harvey partnership
Syncs Walls for AI policies into Harvey's AI platform. This shows the incumbent wall vendor is moving onto AI agents, though it remains focused on legal.
- DeepInspect
A proxy that classifies and blocks AI prompts. Its blog targets MNPI and argues that desk, deal and wall-crossing context should travel with each AI request. It is closest to Mosaic Monitor, but no restricted-list integration or piecing-together test was found.
- RecordPoint RexCommand
Guardrails and baseline drift scoring for Microsoft 365 Copilot and Copilot Studio. It is generic, with no MNPI or restricted-list focus.
- SoftwareOne MNPI SurveillanceShield
AI-driven MNPI surveillance of email, chat and voice that integrates with list-management systems. It does not cover AI agent output, but it is a natural adjacent extension.
- Adversa AI
Continuous AI red teaming for financial services, focused on fraud and security threats. No inside-information or barrier scenarios were found, but it could add them.
- DLA Piper AI Legal Red Teaming
A consulting service, not a product, that red-teams financial firms' AI for legal risk. It could compete on the first assessment engagement.
- Agent Mosaic (arXiv research)
A very recent academic method and benchmark for this exact failure, using synthetic scenarios in cybersecurity, medical and military domains, not finance. The best model leaked in about 7% of ordinary queries, and about 32% of answers still leaked with guardrails. It validates the problem but makes the basic method public.
- MosaicLeaks (arXiv research)
A benchmark of private enterprise documents leaking through a research agent's external web searches. It is related research, not a product.
Checked 2026-10-24
Risks
- The paper uses 30 synthetic examples, so real deal documents may behave differently
- Large banks may build this inside their own red teams
- Bank security review and procurement can take longer than the 3-month horizon
- Copilot-style vendors may add watch-list filters themselves
Strong open space and team fit; tech readiness is the weak point (6).
tech 6 · pull 7 · space 8 · fit 8 · build 7
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
Status history
- 2026-10-24 Flagged as buildable