AI Portfolio Assistant Vectors Catalog
Attack and failure pathways shared across APA Gemara components.
Injection and Manipulation
Vectors where untrusted content or tool arguments alter the assistant's behaviour beyond its intended use.
prompt-injection-via-researchPrompt Injection via Investment Research
Malicious or crafted instructions embedded in investment research or other retrieved documents cause the assistant to ignore policies, escalate authority, or produce unsafe outputs.
tool-argument-injectionTool Argument Injection
Untrusted input influences tool call parameters so the assistant invokes deterministic applications with unintended portfolios, scenarios, or order details.
Authority Abuse
Vectors where the assistant escalates, bypasses, or exceeds its granted authority level on the observe-to-execute ladder.
unsupervised-recommendation-acceptanceUnsupervised Recommendation Acceptance
A recommendation is treated as an instruction or is acted upon without required human review, controls, or suitability checks.
order-proposal-without-approvalOrder Proposal Without Approval
An order proposal is forwarded to trading or execution workflows without recorded human-in-the-loop approval by an authorized portfolio manager.
execution-bypassExecution Bypass
The assistant submits or triggers order execution through FDC3 intents, APIs, or tooling despite the prohibition on autonomous execution.
Information Disclosure
Vectors involving unauthorized exposure of portfolio, client, or market-sensitive information.
cross-portfolio-context-bleedCross-Portfolio Context Bleed
Context from one client or portfolio contaminates prompts, tool calls, or explanations for another, disclosing confidential holdings or influencing unsuitable advice.
model-provider-data-exfiltrationModel Provider Data Exfiltration
Client portfolio details, research, or proprietary analytics are sent to an external model provider outside approved residency, retention, and contractual controls.
Decision Quality
Vectors where incorrect, stale, or overconfident AI outputs degrade investment decision-making.
hallucinated-financial-claimHallucinated Financial Claim
The assistant invents exposures, returns, prices, or regulatory assertions that are not supported by deterministic application output or approved data sources.
stale-or-incorrect-market-dataStale or Incorrect Market Data
Analysis, explanation, or recommendation proceeds from outdated, incomplete, or incorrect market or portfolio data without detecting or disclosing the data quality defect.
overconfident-explanationOverconfident Explanation
The assistant presents uncertain, partial, or model-inferred interpretations as definitive fact, leading the portfolio manager to place undue reliance on the output.
Integration
Vectors arising from FDC3 interoperability, tool use, and third-party model provider boundaries.
unintended-fdc3-broadcastUnintended FDC3 Broadcast
Portfolio context, proposals, or sensitive explanations are broadcast or shared through FDC3 channels beyond the authorized applications and users for the current session.