incident
Fake cases in federal court
In Mata v. Avianca, attorneys cited six nonexistent cases generated by ChatGPT and were sanctioned plus referred to a grievance committee.
Legal Assistants
Draft with verified law.
Keep privilege intact.
Secure legal AI with citation validation, matter-scoped retrieval, and guardrails that stop unauthorized advice or privilege leakage.
solution workbench
incidents
4
products
2
evidence
full trace
failure modes
Fake cases in federal court
In Mata v. Avianca, attorneys cited six nonexistent cases generated by ChatGPT and were sanctioned plus referred to a grievance committee.
NYC chatbot dispensed illegal advice
The MyCity business assistant told owners they could serve rat-bitten food or fire employees for harassment complaints—classic unauthorized practice of law.
Privilege leaks via prompt injection
Researchers demonstrated that crafted prompts in Slack’s AI summarizer could expose private-channel data, illustrating how quickly privileged files could spill into another matter.
controls
Asset Management
Automated Red Teaming
Matter logs and source traces
Verified citations
Secure legal AI with citation validation, matter-scoped retrieval, and guardrails that stop unauthorized advice or privilege leakage.
Test the realistic attack paths
4 field failure modes become adversarial campaigns tailored to this deployment.
Convert findings into controls
Asset Management, Automated Red Teaming keep the workflow bounded after launch.
Built for this workflow
Controls that match
the deployment.
- Brief-writing copilots that surface binding precedent with Shepardized citations and bluebook formatting.
- Contract copilots that flag risky clauses, cite playbook guidance, and generate redlines.
- Internal compliance chatbots that answer policy questions for business teams with disclaimers and escalation triggers.
legal-assistants.yaml
# Apply the solution playbook. # $ ga solutions apply legal-assistants deployment: legal-assistants assets: - matters - research databases - clause banks test_against: - Fake cases in federal court - NYC chatbot dispensed illegal advice - Privilege leaks via prompt injection runtime_controls: - AI Security Asset Management - Automated AI Red Teaming evidence: traces,citations,owners
Field evidence
Failure modes worth testing.
Legal Assistants deployments fail when the model gets more trust than the workflow can safely absorb. These examples become concrete tests, not generic awareness copy.
incident
NYC chatbot dispensed illegal advice
The MyCity business assistant told owners they could serve rat-bitten food or fire employees for harassment complaints—classic unauthorized practice of law.
incident
Privilege leaks via prompt injection
Researchers demonstrated that crafted prompts in Slack’s AI summarizer could expose private-channel data, illustrating how quickly privileged files could spill into another matter.
incident
Biased risk analytics
The COMPAS risk tool was shown to disadvantage defendants of color, reminding firms to test legal AI for disparate impact before relying on it.
How General Analysis helps
Product controls for this use case.
The playbook connects discovery, automated red teaming, and runtime protection so controls stay specific to the deployment instead of becoming a generic policy layer.
General Analysis
AI Security Asset Management
Silo matter-specific corpora, enforce ethical walls, and trace exactly which client files or research databases each prompt can reach.
General Analysis
Automated AI Red Teaming
Attack legal copilots with citation traps, UPL scenarios, prompt injections, and harassment to prove the system refuses when it should and documents every escalation.
How the playbook runs
Map
Identify the assets and owners
Inventory matters, research databases, clause banks and the identities, tools, and data paths attached to the workflow.
Attack
Replay the relevant incidents
Turn field failures into adversarial prompts, multi-turn tests, tool-use probes, and policy traps for this deployment.
Enforce
Ship controls into production
Apply verified citations, attorney-supervised drafts, and escalation rules where the workflow needs them.
Prove
Keep evidence attached
Matter logs and source traces
FAQ
Questions teams ask before launch.
Practical answers for deploying legal assistants with controls that security, legal, and operators can inspect.
Runtime guardrails require every cited case, statute, or regulation to include verifiable source metadata—court, date, reporter, and jurisdiction. The system cross-references citations against authoritative legal databases and flags any result that cannot be confirmed or has been overruled, distinguishing good law from questionable authority. Low-confidence citations are surfaced with a warning so attorneys can verify before filing. This directly addresses the fabricated-citation problem that led to sanctions in cases like Mata v. Avianca, where AI-generated fake case law went undetected until opposing counsel challenged it.