What went wrong elsewhere, and what we built because of it.

What happened

• A solicitor tendered a list and summary of authorities that had been generated with an AI research module in legal software; the citations and summaries were inaccurate and the cases did not exist.
• The output was not reviewed for accuracy before it was handed up. The court referred the matter to the Victorian Legal Services Board and Commissioner.
• On 19 August 2025 the Board varied the practitioner's practising certificate: no practice as a principal and two years of supervised practice.

Sources: Victorian Legal Services Board and Commissioner, 2 September 2025; BN Law, 10 October 2024

The control at issue

Authorities that came from a tool were not checked against a source, and the tool itself produced citations not tied to any document it had read.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  For footnotes, this failure would have been prevented by design in LexIQ: every footnote is rendered by code from the record of a source the system actually retrieved, so a footnote cannot name a case that is not in the sources. Citations the answer makes outside the retrieved evidence are flagged.

• Prevents by design Planned

  An authority verifier that removes from the answer text any authority the retrieved sources do not contain.

• Does not address

  A practitioner who files without reading. No software discharges the duty to check authority against an authoritative source; the courts say so, and every LexIQ answer is labelled a research aid.

What happened

• Submissions filed for an applicant cited cases and quoted a tribunal decision; the Minister's outline said 17 of the cited decisions did not exist.
• The reasons record that the representative had used AI to identify Australian cases and that it gave him case law that did not exist.
• The court directed a referral to the Office of the NSW Legal Services Commissioner.

Sources: Federal Circuit and Family Court of Australia (Division 2), via Wikisource, 31 January 2025

The control at issue

No check of the cited cases against an authoritative source before filing.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  For footnotes, this failure would have been prevented by design in LexIQ: every footnote is rendered by code from the record of a source the system actually retrieved, so a footnote cannot name a case that is not in the sources. Citations the answer makes outside the retrieved evidence are flagged.

• Does not address

  A practitioner who files without reading. No software discharges the duty to check authority against an authoritative source; the courts say so, and every LexIQ answer is labelled a research aid.

What happened

• Submissions contained quotations attributed to a speech in the state parliament that were not real, and citations to Supreme Court decisions that did not exist.
• The judge's associates could not find the cases and asked for copies. The lawyers said they had checked the first citations and assumed the rest were correct.
• The case was delayed by about a day. The judge said AI output must be independently and thoroughly verified before use.

Sources: Associated Press (via WTOP), August 2025

The control at issue

Verification stopped after the first few citations; the opposing party also relied on the submissions without checking.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  For footnotes, this failure would have been prevented by design in LexIQ: every footnote is rendered by code from the record of a source the system actually retrieved, so a footnote cannot name a case that is not in the sources. Citations the answer makes outside the retrieved evidence are flagged.

• Does not address

  Quotations attributed to a parliamentary speech that is not in the sources. LexIQ quotes only the passages it retrieved, but it cannot stop a person pasting text from elsewhere into a submission.

What happened

• The Chief Justice issued a practice note on generative AI and briefed the profession on it on 2 December 2024; the note was amended on 28 January 2025 and took effect on 3 February 2025. Submissions on its review closed on 18 December 2025.
• Among other things it restricts generative AI in affidavits, witness statements and character references, keeps suppressed and other protected material out of uncontrolled tools, and requires practitioners to verify citations themselves.
• The Federal Court's practice note of 16 April 2026 and the profession's joint statement of December 2024 point the same way.

Sources: Supreme Court of New South Wales, amended 28 January 2025, in effect 3 February 2025; page updated 23 March 2026; Federal Court of Australia, 16 April 2026; Law Society of NSW, Legal Practice Board of WA, Victorian Legal Services Board and Commissioner, 6 December 2024

The control at issue

The note responds to the risks courts had seen: invented authority, and confidential material entered into tools outside the parties' control.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  Protected material sent to an uncontrolled tool: at the Maximum level, or in a matter set to block, nothing leaves the firm's server, and the code refuses the call.

• Does not address Planned

  A rule against AI-generated content in affidavits and witness statements is for the product to respect, not to solve. A guard that declines such drafting and a disclosure helper are planned.

What happened

• Two cases referred to the court's jurisdiction over lawyers' duties. In one, fictitious cases were put before the court; in the other, 45 citations were relied on and in 18 the case cited did not exist.
• The court referred lawyers to their regulators and decided not to begin contempt proceedings, while warning that this was not a precedent.
• The court said the critical safeguard is to check any output against an authoritative source.

Sources: High Court of Justice, King's Bench Division, Divisional Court, 6 June 2025

The control at issue

Lawyers relied on research they had not checked, including research done by a lay client.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  For footnotes, this failure would have been prevented by design in LexIQ: every footnote is rendered by code from the record of a source the system actually retrieved, so a footnote cannot name a case that is not in the sources. Citations the answer makes outside the retrieved evidence are flagged.

• Prevents by design Planned

  An authority verifier that removes from the answer text any authority the retrieved sources do not contain.

• Does not address

  A practitioner who files without reading. No software discharges the duty to check authority against an authoritative source; the courts say so, and every LexIQ answer is labelled a research aid.

What happened

• A motion drafted with a general-purpose AI chatbot cited cases, quotations and internal citations that did not exist. Neither the other side nor the court could find them.
• The court imposed a USD 5,000 sanction on the lawyers responsible.

Sources: Wikipedia summary of the decision, sanctions opinion of 22 June 2023; High Court of Justice, King's Bench Division, Divisional Court, 6 June 2025

The control at issue

Citations were taken from a chat tool with no link to any source document.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  For footnotes, this failure would have been prevented by design in LexIQ: every footnote is rendered by code from the record of a source the system actually retrieved, so a footnote cannot name a case that is not in the sources. Citations the answer makes outside the retrieved evidence are flagged.

• Does not address

  A practitioner who files without reading. No software discharges the duty to check authority against an authoritative source; the courts say so, and every LexIQ answer is labelled a research aid.

What happened

• On 28 April 2023 the firm learnt that a ransomware group claimed to have taken data from it. The firm says the attacker took information from a confined part of its system, and some of it was published on the group's forum for three weeks in June 2023.
• The National Cyber Security Coordinator said sensitive personal and government information from a number of Australian Government entities was released. The firm obtained an injunction from the NSW Supreme Court restraining further publication.
• The OAIC opened an investigation into the firm's security and its notification of affected people; it has said it will not comment until the investigation ends.

Sources: HWL Ebsworth Lawyers, undated statement; AAP, July 2023; Office of the Australian Information Commissioner, 21 February 2024

The control at issue

Public sources do not say how the attacker got in, and we make no suggestion about it. The general lesson is that a firm's document store holds many clients' material in one place.

How this offering is designed differently

• Contains or detects In development

  On our platform each firm has its own database, and client text is sealed with that firm's own key, so taking a copy of the data does not by itself reveal client text. In a comparable attack on the platform, exposure would have been limited by these measures; this is a design property, not a claim about what would have happened at any firm.

• Contains or detects In development

  AegisIQ's exfiltration signals are built and measured against replayed attacks; running them live comes next.

• Does not address

  A compromised computer on the firm's own network, or documents kept outside the platform, are beyond what the platform can protect.

What happened

• The company reported a cyber security incident after its servers were compromised on 22 November 2020; attackers posted screenshots and threatened to publish data unless paid.
• The company said at the time that it had seen no evidence of data exfiltration and nothing indicating its customers' networks were compromised.

Sources: iTWire, 23 November 2020

The control at issue

A supplier holding many firms' documents became a single point of exposure.

How this offering is designed differently

• Contains or detects Built, in alpha testing

  Each firm's data sits in its own database and a request that cannot be tied to the right firm is refused, which limits how far one compromise reaches.

• Contains or detects Built, in alpha testing

  Encrypted LexIQ backups to two recipients, with no unencrypted copy on disk, keep recovery possible without paying anyone; the other services follow.

What happened

• Malware hit the firm's systems and it shut down digital operations in offices worldwide; two days later email had not been fully restored.
• The firm said it had seen no evidence that client data was taken.

Sources: Fortune, 29 June 2017

The control at issue

One network shared across offices let destructive malware reach everything at once.

How this offering is designed differently

• Contains or detects Built, in alpha testing

  Services authenticate each other with mutual TLS and short-lived tokens, so no service trusts a request because of where it came from, and each layer can run on its own host.

• Contains or detects Built, in alpha testing

  Encrypted LexIQ backups with weekly restore drills make recovery a rehearsed step. Backups that cannot be altered, on separate infrastructure, are still open.

• Does not address

  Malware on a firm's own desktops and file servers is outside the platform.

What happened

• More than 11.5 million of the firm's financial and legal records were leaked and published by journalists.
• When it announced its closure, the firm blamed reputational damage and said it had been the victim of a cyber-attack. How the records left the firm has not been established in a source we could verify.

Sources: International Consortium of Investigative Journalists, 3 April 2016; OCCRP, 15 March 2018

The control at issue

Records in bulk left the firm without being noticed in time.

How this offering is designed differently

• Contains or detects In development

  Bulk export is the kind of event AegisIQ's exfiltration signals are built to notice; they are not yet live. They are measured against replayed attacks.

• Does not address

  A person with legitimate access who decides to disclose records. The access audit records who exported what, but software cannot remove that risk.

What happened

• The company detected unauthorised activity, confirmed as ransomware, and took its systems offline globally to contain it.
• It said it had no evidence of unauthorised transfer or misuse of data.

Sources: ABA Journal, 3 March 2020

The control at issue

When a hosted legal platform goes offline, every client loses access at once.

How this offering is designed differently

• Contains or detects Planned

  Hybrid two-site operation keeps a firm's on-premises site working when the cloud site is down, and catches up afterwards. Proven in a simulator so far; live sync is planned for December 2026.

• Contains or detects Built, in alpha testing

  Each layer has a defined degraded behaviour, so search can keep working when another part is down.

What happened

• In Dayal, the list of non-existent authorities came from an AI research module embedded in the practitioner's legal software, not from a separate chat tool.

Sources: BN Law, 10 October 2024

The control at issue

A tool inside trusted practice software produced output that looked like the rest of the system's records.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  In LexIQ every answer is labelled a generated research aid, and every footnote is rendered from the record of a retrieved source with its pinpoint, so the practitioner can open what was cited.

• Does not address

  A practitioner who files without reading. No software discharges the duty to check authority against an authoritative source; the courts say so, and every LexIQ answer is labelled a research aid.

What happened

• In copyright litigation, a United States court ordered OpenAI to preserve output logs that would otherwise have been deleted, including deleted and temporary chats and data from API customers without a zero data retention agreement.
• Enterprise and education customers, and API customers with zero data retention, were not affected. OpenAI opposed the order; an order of 9 October 2025 ended the general obligation, with some exceptions.

Sources: Simon Willison's weblog, 5 June 2025, updated 22 October 2025

The control at issue

Customers of a hosted provider found their data held longer than the provider's stated retention, because of litigation they were not party to.

How this offering is designed differently

• Prevents by design Built, in alpha testing

  At the Maximum level, and in matters set to block, no question or document reaches a hosted provider, so no provider can be ordered to keep it.

• Does not address

  At the Standard and Enhanced levels, a question and the passages needed to answer it do go to a hosted provider outside Australia under its retention terms. Firms see a warning naming the region and retention before they choose those levels.