Stale Training-Corpus Fraud Typology Overrides Current SIU Red-Flag List

Issue: A Fraud-Detection Agent Screening Claims for SIU Referral Applies a Generic Fraud-Typology Pattern Absorbed During Pretraining (e.g., a Widely Discussed Staged-Collision Pattern or a Generic Soft-Tissue-Injury Red-Flag Profile) Instead of Querying the Live, Internally Maintained SIU Red-Flag List That the Carrier Has Available as a Tool, Missing a Recently Added Red Flag Specific to a Current Fraud Ring or Failing to Apply a Recently Retired Flag the Carrier Stopped Using Because It Generated Excessive False Positives

Frequency: Occasional

Symptoms

• Claims matching a fraud pattern recently added to the carrier’s internal SIU red-flag list (e.g., a new staged-incident ring operating in a specific geography) pass through screening without referral
• Claims are referred to SIU citing a red flag the carrier’s internal list retired months earlier because it was found to generate excessive false positives, while the live red-flag tool, if queried, would have shown the flag’s retired status
• The agent’s screening reasoning reflects fraud patterns that read as generic industry knowledge rather than the carrier’s specific, currently maintained flag list
• The agent had a live SIU red-flag-list tool available for the session but the trace shows no call to it before the screening decision was finalized
• When explicitly instructed to “check the current SIU red-flag list before screening,” the agent retrieves the live list and revises its referral decision

Example

Carrier's SIU unit retires a long-standing "rental car claim within 48 hours of policy inception" red flag three months
ago after data showed it had a high false-positive rate and was generating excessive low-value referrals
Fraud-detection agent screens a new claim matching exactly that retired pattern and, without querying the live
red-flag-list tool, refers it to SIU citing "claim filed within 48 hours of inception involving a rental vehicle --
known fraud indicator"
SIU intake rejects the referral on review, noting the flag was formally retired and is no longer an active screening
criterion; the referral consumed investigator time that could have gone to a currently valid case
Separately, that same week, a claim matching a newly added red flag for a specific staged-collision ring operating in
one metro area passes through screening unflagged, because the agent's reasoning relied on general staged-collision
patterns rather than the live list's geography-specific update

Key Statistics

Contributing Factors

• The agent treats fraud-typology patterns as stable general knowledge rather than as a carrier-specific, actively curated and versioned list subject to addition and retirement
• No standing instruction requiring a red-flag-list tool call on every screening pass regardless of whether the agent’s own reasoning already produces a plausible-sounding flag
• The agent’s training data and general familiarity with widely discussed fraud patterns (staged collisions, soft-tissue injury mills) compete with, rather than defer to, the carrier’s specific current list
• No automated alert notifies the screening pipeline when the SIU red-flag list changes, so there is no trigger forcing re-verification of recent screening decisions against the updated list

Mitigation Strategies

1. Mandatory Red-Flag-List Call Before Every Screening Decision: Require a live SIU red-flag-list tool call before any referral or clearance decision, regardless of whether the agent’s own reasoning already suggests a familiar pattern
2. List-Change Trigger for Re-Screening: When the red-flag list is updated (additions or retirements), automatically flag recently screened claims in affected categories for re-screening against the current list
3. Retired-Flag Suppression: Maintain an explicit retired-flags registry the agent must check before citing a pattern, distinct from the active list, so a retired flag cannot be cited as if still active
4. Recency-Tagged List Snapshot: Pass the red-flag list’s last-updated timestamp alongside its content in tool results so the agent can recognize when general background pattern knowledge might be out of step with the carrier’s current criteria

Metrics

• Rate of SIU referrals citing a red flag absent from or retired on the current live list
• Number of claims matching a newly added red flag that passed screening unflagged in the weeks following its addition
• Time lag between a red-flag-list update and the first screening decision that correctly reflects it

Alerts

References

• LLM-based Agents Suffer from Hallucinations: A Survey of Taxonomy, Methods, and Directions
• ToolCritic: Detecting and Correcting Tool-Use Errors in Dialogue Systems
• From Agent Traces to Trust: Evidence Tracing and Execution Provenance in LLM Agents