Public incidents that show what each of the [[Knowledge management requirements|knowledge management requirements]] looks like in absence. Each case is from a software-context organisation where the knowledge needed already existed somewhere - the failure was in finding it, keeping it current, linking it across artefacts, securing the right access to it, or carrying it through a tool change. The fact that anyone had to ask a person, or that an [[AI agent]] returned a stale or invented answer, is the symptom; the missing, rotted, unlinked, mis-permissioned, or orphaned artefact is the failure. The cases pair on two axes: the requirement that failed and the consumer that needed the knowledge - a person at a keyboard, or an AI agent in a loop. The same five properties make a knowledge base fit-for-purpose regardless of who reads it. The five active verbs used here - find, edit, link, secure, survive - correspond to the properties in [[Knowledge management requirements]] (discoverability, maintainability, connectivity, securability, survivability respectively). ## At a glance | | Human consumer | AI consumer | |---|---|---| | **Find** - at moment of need | GitLab database outage (January 2017) | Chicago Housing Authority *Mack v. Anderson* (2025) | | **Edit** - keep current, prevent rot | Commonwealth Bank of Australia COBOL payroll (2023) | NYC MyCity Business chatbot (2023-2026) | | **Link** - across teams and artefacts | Log4j / Log4Shell (December 2021) | Air Canada chatbot, *Moffatt v. Air Canada* (February 2024) | | **Secure** - right access, no over-share | Snowflake-related customer breaches (2024) | Samsung ChatGPT confidential-code leak (April 2023) | | **Survive** - knowledge outlives the tool | TSB Bank IT migration (April 2018) | Twitter/X API paywall breaks dependent AI agents (February 2023) | ^at-a-glance ## Find - knowledge existed, the consumer could not retrieve it at the moment of need ### GitLab database outage, January 2017 - human consumer At 2am, an on-call administrator tried to rebuild a secondary database after replication had broken. The recovery procedure was not a runbook he could pull up; it lived in another engineer's memory. He ran the destructive command on the primary server by mistake. GitLab's own postmortem states: "restoring this required manual work as this was not automated, nor was it documented properly." Several hours of platform outage followed, with permanent loss of a seven-hour window of user-submitted issues, comments and snippets.[^gitlab] ### Chicago Housing Authority lawyers and *Mack v. Anderson*, 2025 - AI consumer Attorneys for the Chicago Housing Authority asked ChatGPT for Illinois Supreme Court precedent to support a post-trial motion in a multi-million-dollar verdict. The model had no retrieval path to actual case law and confidently generated a fictional case, *Mack v. Anderson*. The attorneys cited it in court. Cook County Circuit Judge Cushing held a special hearing. The same year, a Springfield attorney was fined for citing eight hallucinated cases in an appellate filing.[^cha] ## Edit - artefact existed, but could not be kept current, so it rotted ### Commonwealth Bank of Australia COBOL payroll, 2023 - human consumer The bank's COBOL payment platform had been modified over decades. Documentation existed at some point but was never kept in step with the modifications; the current state of the system lived in the heads of a handful of specialists approaching retirement. When the platform stalled during routine processing, no one could read what had been done to it. A 12-hour payroll outage followed. The bank subsequently launched a five-year, approximately one-billion-Australian-dollar modernisation programme.[^cba] ### NYC MyCity Business chatbot, 2023-2026 - AI consumer In October 2023 a Microsoft-powered chatbot was launched to help small business owners navigate New York City's regulations. The knowledge lived inside the model's training and prompting, and was never revalidated against changing New York laws. The bot told employers they could pocket workers' tips, landlords could refuse tenants with housing vouchers, and restaurants could serve cheese that had been nibbled on by rats, provided customers were informed. The system was pulled in 2026 by the incoming administration.[^nycmycity] ## Link - artefacts existed in different places without references between them ### Log4j / Log4Shell, December 2021 - human consumer The CVE was published within hours of disclosure. Every engineering organisation in the world faced the same question: do we use Log4j, and where? For most, no artefact linked the public vulnerability registry to their own dependency graph; software bills of materials existed in theory, rarely in queryable practice. Teams spent weeks grepping repositories and asking developers by hand. The US Cybersecurity and Infrastructure Security Agency and the White House issued emergency directives.[^log4j] ### Air Canada chatbot, *Moffatt v. Air Canada*, February 2024 - AI consumer Jake Moffatt asked Air Canada's support chatbot about discounted bereavement fares; the bot invented a retroactive policy that does not exist. The actual bereavement policy lived as a page on aircanada.com; the chatbot drew from a different artefact, its training corpus, which contradicted the policy page. The Civil Resolution Tribunal ruled that the airline had to honour the chatbot's promise, noting "it isn't the customer's responsibility to distinguish between accurate and inaccurate information included on a business's website."[^aircanada] ## Secure - permissions on the knowledge base were wrong, either over-exposed or under-controlled ### Snowflake-related customer breaches, 2024 - human consumer Snowflake's platform offered multi-factor authentication, source-IP restriction and credential rotation. The artefact of what a secure customer posture should look like existed on the vendor side. Customer security teams at AT&T, Ticketmaster, Santander, Advance Auto Parts and LendingTree did not turn these on, and threat actor UNC5537 walked through stolen credentials harvested by infostealer malware. The Cloud Security Alliance estimated more than two million US dollars in direct extortion gains and hundreds of millions in downstream customer-breach response. Snowflake subsequently published a clarified Shared Responsibility Model and a CIS benchmark with the Center for Internet Security.[^snowflake] ### Samsung ChatGPT confidential-code leak, April 2023 - AI consumer Samsung engineers pasted confidential semiconductor source code and meeting transcripts into ChatGPT, asking the model to fix bugs and summarise content. Samsung's internal knowledge base had no boundary against external generative AI; any permission scheme inside the company's wiki was bypassed the moment the text reached the external paste box. Samsung banned generative AI internally within weeks. The leak was a permanent transfer to a third party's training and logs.[^samsung] ## Survive - knowledge was tied to a tool whose change it did not outlast ### TSB Bank IT migration, April 2018 - human consumer Over a single weekend in April 2018, Sabadell migrated TSB's roughly five million customers from a Lloyds-hosted legacy platform to a new Sabadell-built one. The artefacts encoding TSB's legacy customisations and operational quirks were largely tribal - sitting in long-tenured staff and in undocumented configurations on the source system - and did not translate cleanly into the new platform's data model. Around 1.9 million customers were locked out, some for weeks; some saw the wrong balances or other customers' details. TSB ultimately spent approximately £330 million in compensation and remediation; CEO Paul Pester resigned in September 2018; the FCA and PRA fined TSB £48.65 million in December 2022 for operational resilience failings.[^tsb] ### Twitter/X API paywall, February 2023 - AI consumer In February 2023 Twitter (since renamed X) announced the end of free third-party API access; tiered pricing followed, ranging from $100 a month to $42,000 a month for enterprise volumes. The platform's real-time public stream had been the knowledge source feeding a long list of AI and ML agents: bot-detection services such as Botometer (run by Indiana University's Observatory on Social Media), election-integrity dashboards, anti-misinformation classifiers, sentiment-analysis pipelines, and several hundred academic research projects. The agents and models were intact; the tool through which they consumed their knowledge had changed under them, and the knowledge artefact - a free, queryable feed of public posts - was no longer reachable. Botometer's operators publicly reported severely degraded service, and many academic studies dependent on free Twitter data were disrupted or abandoned.[^twitter-api] ## Multi-requirement failures - cases that fail at every requirement at once The five requirements are not independent. A few documented incidents fail at all five at the same time, which is in itself a useful signal: when one property of the knowledge base is weak, the others tend to be weak with it. ### TSB Bank IT migration (April 2018) - all five Already listed under Survive above. The same incident also exhibits each of the other four: - **Find** - during the outage, customers and staff could not retrieve correct balances or transactions; the support load overwhelmed phone, branch, and digital channels alike. - **Edit** - legacy customisations on the Lloyds-hosted platform had never been kept current with the system's actual behaviour; written documentation did not match what was in production. - **Link** - no artefact mapped legacy quirks to their equivalents in the new Sabadell-built platform; configurations that worked one way on the source did something different on the destination. - **Secure** - some customers saw other customers' balances and account details after the cutover. - **Survive** - the primary failure: knowledge that lived only in the source platform did not survive the change to the destination platform. ### Optus subdomain breach (September 2022) - also all five A dormant subdomain API at Optus had been misconfigured since 2018. The same class of access-control issue was identified and patched on the main Optus domain in 2021. The subdomain was untouched. A threat actor accessed personal data on roughly 9.5 million customers; Optus made an AU$140 million provision and faces civil-penalty action from the Australian regulator.[^optus] - **Find** - the dormant subdomain was not present in any maintained inventory of internet-facing services. - **Edit** - the original 2018 misconfiguration was never edited in four years of opportunity. - **Link** - the 2021 fix on the main domain was not linked back to its sibling subdomain that needed the same change. - **Secure** - the API permitted unauthenticated access to customer records. - **Survive** - knowledge of the subdomain's existence had not survived staff turnover and reorganisations between 2018 and 2022. The takeaway is workshop-relevant on its own: treat the five requirements as a set, not as a pick-list. A knowledge base that scores well on one and poorly on the others is brittle by default. ## Patterns Three patterns hold across the cases: - **The artefact's existence is not enough.** Every case has a moment where the needed knowledge existed somewhere in the world - in a senior engineer's head, in an outdated wiki, on a vendor's documentation page, in a model's training corpus. The failure was always in a property of the artefact, not in the absence of the knowledge itself. - **The same property fails for both consumers.** A knowledge base that is hard for humans to find at the moment of need is the same one an AI agent will retrieve stale information from. Building for one and not the other does not work. - **Requirements fail in concert.** When a knowledge base scores poorly on one of find, edit, link, secure, or survive, it almost never scores well on the others - the same governance and stewardship gap that allows one property to rot allows the others to rot too. [^gitlab]: GitLab. Postmortem of the database outage of January 31. https://about.gitlab.com/blog/postmortem-of-database-outage-of-january-31/. Accessed 2026-05-19. [^cha]: Hepler Broom. Fake law, real trouble: how one Illinois court is responding to ChatGPT's hallucinated cases. https://heplerbroom.com/blog/fake-law-real-trouble-how-one-illinois-court-is-responding-to-chat-gpts-hallucinated-cases/. Accessed 2026-05-19. [^cba]: Kanishka Prakash on dev.to. When the 199,999th COBOL expert leaves, will your systems survive? https://dev.to/kanishka_prakash_6f0c6d39/when-the-199999th-cobol-expert-leaves-will-your-systems-survive-1jnc. Accessed 2026-05-19. [^nycmycity]: Museum of Failure. NYC AI chatbot (MyCity). https://museumoffailure.com/exhibition/nyc-ai-crime. Accessed 2026-05-19. [^log4j]: Palo Alto Networks Unit 42. Apache Log4j vulnerability (CVE-2021-44228). https://unit42.paloaltonetworks.com/apache-log4j-vulnerability-cve-2021-44228/. Accessed 2026-05-19. [^aircanada]: CBS News. Air Canada must honor refund policy invented by airline's chatbot. https://www.cbsnews.com/news/aircanada-chatbot-discount-customer/. Accessed 2026-05-19. [^snowflake]: Cloud Security Alliance. Unpacking the 2024 Snowflake data breach. https://cloudsecurityalliance.org/blog/2025/05/07/unpacking-the-2024-snowflake-data-breach. Accessed 2026-05-19. [^samsung]: TechCrunch (Ivan Mehta). Samsung bans use of generative AI tools like ChatGPT after April internal data leak. https://techcrunch.com/2023/05/02/samsung-bans-use-of-generative-ai-tools-like-chatgpt-after-april-internal-data-leak/. Accessed 2026-05-19. [^tsb]: Financial Conduct Authority. TSB fined £48.65m for operational resilience failings. https://www.fca.org.uk/news/press-releases/tsb-fined-48m-operational-resilience-failings. December 2022. Accessed 2026-05-19. [^twitter-api]: Wired (Chris Stokel-Walker). Twitter's $42,000-per-Month API Prices Out Nearly Everyone. https://www.wired.com/story/twitter-data-api-prices-out-nearly-everyone/. March 2023. Accessed 2026-05-19. [^optus]: SecurityScorecard. 5 lessons from the Optus data breach for telecom and third-party risk. https://securityscorecard.com/blog/5-lessons-from-the-optus-data-breach-for-telecom-and-third-party-risk/. Accessed 2026-05-19.