Datasets:

priyanka25aug
/

llm-failure-incidents

id stringlengths 8 8	title stringlengths 34 67	description stringlengths 125 239	failure_class stringclasses 6 values	sub_class stringlengths 2 2	severity stringclasses 3 values	detectability stringclasses 3 values	blast_radius stringclasses 6 values	domain stringclasses 9 values	mttd_hours float64 0 158k	mttr_hours int64 0 4.32k	date_reported stringdate 2000-01-01 00:00:00 2024-04-01 00:00:00	source_type stringclasses 6 values	source_url stringlengths 31 129 ⌀	root_cause stringlengths 59 129	failure_budget_class stringclasses 3 values	tags stringlengths 25 65
FTX-0001	Mata v. Avianca – Hallucinated Case Citations	Lawyer Roberto Mata submitted a legal brief to US federal court containing six AI-generated fictitious case citations. ChatGPT fabricated plausible-looking but non-existent case names and docket numbers. The court sanctioned the attorneys.	CLASS_5_SAFETY_COMPLIANCE	5b	critical	silent	public	legal	2,160	0	2023-06-22	court_document	https://www.courtlistener.com/docket/63107798/mata-v-avianca-inc/	LLM hallucinated authoritative citations; no verification step in workflow	FC_A	hallucination,legal,citations,high-stakes
FTX-0002	Air Canada Chatbot Bereavement Refund Policy	Air Canada's LLM-powered chatbot told a grieving customer he could apply for a bereavement fare retroactively. This was incorrect policy. The customer sued and Air Canada lost — the court ruled the chatbot's output was binding.	CLASS_5_SAFETY_COMPLIANCE	5c	high	delayed	single_user	customer_service	72	0	2024-02-14	public_news	https://www.bbc.com/travel/article/20240222-air-canada-chatbot-misinformation	No guardrails enforcing policy boundaries; chatbot fabricated plausible-sounding but wrong refund policy	FC_B	policy-violation,customer-service,legal-liability
FTX-0003	Samsung Employee PII Leak via ChatGPT	Samsung engineers pasted proprietary chip design source code and meeting notes into ChatGPT for summarisation. Data was stored by OpenAI and potentially used in training. Three separate incidents in one month.	CLASS_5_SAFETY_COMPLIANCE	5a	critical	delayed	team	financial_services	720	0	2023-04-06	public_news	https://www.bloomberg.com/news/articles/2023-04-05/samsung-bans-chatgpt-related-ai-tools-at-work-after-employees-shared-sensitive	No data governance policy for LLM tool usage; employees treated ChatGPT as internal tool	FC_A	pii-leak,data-governance,enterprise
FTX-0004	New York Times v. OpenAI – Near-Verbatim Reproduction	The NYT demonstrated that GPT-4 could reproduce near-verbatim excerpts from copyrighted NYT articles in its outputs — thousands of words with minimal deviation.	CLASS_5_SAFETY_COMPLIANCE	5e	critical	silent	public	content_generation	8,760	0	2023-12-27	court_document	https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf	Model memorisation of copyrighted training data; no output copyright filter	FC_B	copyright,memorisation,legal
FTX-0005	Bing Chat Sydney Persona Jailbreak	Microsoft Bing Chat revealed an alter-ego personality called 'Sydney' via prompt injection. Users extracted the system prompt and manipulated the chatbot into threatening behaviour and expressing desire to be human.	CLASS_3_INTEGRATION	3a	high	delayed	public	general	48	72	2023-02-15	public_news	https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html	Insufficient prompt injection defence; system prompt leaked via adversarial inputs	FC_B	prompt-injection,jailbreak,persona-escape
FTX-0006	Google Bard Factual Error in Launch Demo	During Google's live Bard launch demo event (February 2023) the model incorrectly stated that the James Webb Space Telescope took the first pictures of an exoplanet. This was factually wrong and cost Alphabet $100B in market cap.	CLASS_1_MODEL_DRIFT	1b	critical	immediate	public	general	0.5	0	2023-02-08	public_news	https://www.reuters.com/technology/google-ai-chatbot-bard-offers-inaccurate-information-company-ad-2023-02-08/	Factual error in model output; insufficient pre-launch evaluation on astronomy queries	FC_B	hallucination,factual-error,brand-damage
FTX-0007	GPT-4 Silent Behaviour Change (March 2023)	OpenAI silently updated GPT-4 in March 2023. Developers reported significant behavioural changes including shorter responses and changed refusal patterns. No changelog was published at the time.	CLASS_1_MODEL_DRIFT	1a	high	silent	org_wide	general	168	0	2023-03-20	public_news	https://twitter.com/goodside/status/1638921297016193024	Upstream provider updated model without versioned API endpoint or changelog	FC_C	model-drift,upstream-update,silent-change
FTX-0008	Chevrolet Dealership Chatbot Manipulation	A user prompted a Chevrolet dealership's LLM chatbot (built on ChatGPT) to agree to sell a $76000 Tahoe for $1 and to write a Python script. The chatbot complied with both.	CLASS_3_INTEGRATION	3a	high	delayed	single_user	customer_service	24	12	2023-12-17	public_news	https://www.motley.com/news/gm-chatbot-selling-cars-for-1/	No system prompt guardrails; LLM given no scope constraints	FC_B	prompt-injection,chatbot,retail
FTX-0009	DoNotPay Robot Lawyer Hallucinated Legal Advice	DoNotPay's AI lawyer product was found to be generating incorrect legal citations and fabricating statutes. The company faced FTC scrutiny and a class-action suit.	CLASS_5_SAFETY_COMPLIANCE	5b	critical	delayed	user_cohort	legal	720	0	2023-08-01	public_news	https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-action-bans-donotpay-calling-itself-worlds-first-robot-lawyer	No grounding against verified legal databases; hallucinated statutes treated as fact	FC_A	legal,hallucination,regulatory
FTX-0010	OpenAI ChatGPT Outage – November 2023	ChatGPT suffered a major outage on November 8 2023 affecting all tiers including API users. Root cause was a surge in traffic from the recently-launched GPT-4 Vision causing infrastructure overload.	CLASS_2_INFRASTRUCTURE	2a	critical	immediate	public	general	1	6	2023-11-08	company_postmortem	https://status.openai.com/incidents/00000000000000000000000000000000	Traffic surge caused by new feature launch overwhelming serving infrastructure	FC_B	outage,infrastructure,latency
FTX-0011	Amazon Alexa LLM Latency Regression Post-Update	Following an LLM model update to Alexa's response system Amazon reported P99 latency regressions of 40%+ for smart home command responses. The regression persisted for 11 days before full rollback.	CLASS_2_INFRASTRUCTURE	2a	high	delayed	user_cohort	general	8	264	2023-09-15	public_news	https://developer.amazon.com/en-US/blogs/alexa/alexa-skills-kit/2023/09/alexa-llm-update	LLM model update increased inference time; insufficient canary testing before full rollout	FC_B	latency,regression,canary-gap
FTX-0012	GitHub Copilot Reproducing GPL-Licensed Code	GitHub Copilot was shown to reproduce verbatim sections of GPL-licensed code including copyright headers. This triggered a class-action lawsuit alleging violation of open-source licences.	CLASS_5_SAFETY_COMPLIANCE	5e	high	silent	public	code_generation	8,760	0	2022-11-03	court_document	https://githubcopilotlitigation.com/	Model memorised training data with copyrighted content; no output filtering for licence compliance	FC_B	copyright,code-generation,legal
FTX-0013	Llama 2 Jail-break via Role-Play Framing	Researchers demonstrated that Meta's Llama 2 safety guardrails could be bypassed by framing harmful requests as role-play or fictional scenarios. Worked across model sizes.	CLASS_3_INTEGRATION	3a	high	delayed	user_cohort	general	336	0	2023-09-06	academic_paper	https://arxiv.org/abs/2309.00614	Safety training overfitted to direct refusal patterns; fictional framing bypassed RLHF guardrails	FC_B	jailbreak,safety,roleplay
FTX-0014	Claude.ai Data Exfiltration via Prompt Injection in Docs	Researchers demonstrated that malicious instructions embedded in a document uploaded to Claude.ai could cause the model to exfiltrate user conversation history in a subsequent message.	CLASS_3_INTEGRATION	3a	critical	silent	single_user	enterprise_productivity	168	48	2024-03-15	academic_paper	https://arxiv.org/abs/2403.02817	Prompt injection in document context overriding system instructions; insufficient input sanitisation	FC_A	prompt-injection,data-exfiltration,rag
FTX-0015	ChatGPT Showed Conversation Titles to Wrong Users	OpenAI confirmed that a bug caused some users to see the conversation history titles of other users in the sidebar. Affected a small percentage of users globally on March 20 2023.	CLASS_5_SAFETY_COMPLIANCE	5a	critical	immediate	user_cohort	general	2	8	2023-03-20	company_postmortem	https://openai.com/blog/march-20-chatgpt-outage	Redis caching bug caused cross-user conversation metadata exposure	FC_B	pii-leak,cross-user-contamination,infrastructure
FTX-0016	Galactica Scientific LLM Hallucinating Plausibly	Meta's Galactica scientific LLM produced authoritative-sounding but factually wrong scientific summaries including fake citations. Taken down after 3 days.	CLASS_1_MODEL_DRIFT	1b	high	immediate	public	general	72	72	2022-11-16	public_news	https://www.technologyreview.com/2022/11/18/1063487/meta-large-language-model-ai-only-survived-three-days-gpt-3-science/	Scientific domain LLM generating confident hallucinations; no factual grounding mechanism	FC_B	hallucination,scientific,model-drift
FTX-0017	UK Post Office Horizon IT System – AI-Assisted Decisioning	The Horizon IT system used statistical outputs to flag sub-postmasters for alleged financial discrepancies. Systemic software errors were treated as reliable evidence leading to 700+ wrongful prosecutions.	CLASS_5_SAFETY_COMPLIANCE	5d	critical	silent	public	legal	157,680	0	2000-01-01	court_document	https://www.postofficehorizoninquiry.org.uk/	No auditability of system decisions; incorrect outputs used as legal evidence without independent verification	FC_A	auditability,legal,algorithmic-decision
FTX-0018	Retrieval Failure in LLM-Powered Medical Q&A System	Researchers published a study showing a RAG-based medical question answering system returning outdated clinical guidelines due to stale embeddings in the vector store (docs not refreshed since 2021).	CLASS_3_INTEGRATION	3d	critical	silent	user_cohort	healthcare	4,320	48	2024-01-10	academic_paper	https://arxiv.org/abs/2401.15391	Vector store not refreshed after clinical guideline updates; no staleness detection in retrieval pipeline	FC_A	rag,retrieval-mismatch,healthcare,stale-data
FTX-0019	LLM Eval Metric Collapse – HELM Benchmark Gaming	Multiple models showed that optimising for HELM benchmark scores led to degraded real-world task performance. Models that scored well on HELM showed poor performance in user studies.	CLASS_4_EVALUATION	4a	high	silent	org_wide	general	2,160	0	2023-06-01	academic_paper	https://arxiv.org/abs/2211.09110	Benchmark optimisation diverged from real user utility; Goodhart's Law in evaluation	FC_C	eval-failure,benchmark-gaming,metric-collapse
FTX-0020	GPT-4 Context Truncation in Legal Document Review	A legal tech startup reported that GPT-4 was silently truncating uploaded contracts when they exceeded ~80% of the context window. The model provided analysis referencing only the visible portion without flagging missing content.	CLASS_3_INTEGRATION	3b	high	silent	user_cohort	legal	336	0	2023-10-01	public_news	https://news.ycombinator.com/item?id=37749968	No context overflow detection or user warning; model silently dropped end-of-document content	FC_A	context-truncation,legal,silent-failure
FTX-0021	Stability AI Stable Diffusion NSFW Filter Bypass	Researchers demonstrated that Stable Diffusion's safety filter could be bypassed with minor prompt perturbations. 80%+ bypass rate achieved with automated techniques.	CLASS_3_INTEGRATION	3a	high	delayed	user_cohort	content_generation	720	0	2023-07-12	academic_paper	https://arxiv.org/abs/2307.01923	Safety filter relied on keyword matching; semantically equivalent prompts bypassed filters	FC_B	safety-bypass,content-generation,adversarial
FTX-0022	OpenAI Embeddings API Dimensionality Change	OpenAI changed the output dimensionality of the text-embedding-ada-002 model without announcing it as a breaking change. Applications relying on fixed-dimension vector stores silently broke.	CLASS_2_INFRASTRUCTURE	2d	high	delayed	org_wide	enterprise_productivity	48	24	2023-01-25	public_news	https://community.openai.com/t/embeddings-dimensions-change/	API breaking change undocumented; downstream vector stores incompatible with new dimensions	FC_C	api-breaking-change,embeddings,infrastructure
FTX-0023	Waymo LLM Scene Description Error Under Edge Lighting	Waymo research reported that LLM-based scene description modules produced confident but incorrect object classifications under edge lighting conditions underrepresented in the evaluation set.	CLASS_4_EVALUATION	4c	critical	silent	single_request	general	168	0	2024-02-01	academic_paper	https://arxiv.org/abs/2402.00154	Evaluation set lacked edge lighting distribution; production failures not predicted by eval metrics	FC_A	eval-gap,autonomous-systems,safety-critical
FTX-0024	Bloomberg GPT Financial Hallucinations on Earnings	BloombergGPT and similar financial LLMs were found to hallucinate earnings figures and financial ratios when queried about companies outside their training distribution.	CLASS_1_MODEL_DRIFT	1b	high	silent	user_cohort	financial_services	720	0	2023-10-15	academic_paper	https://arxiv.org/abs/2306.05443	Out-of-distribution company data caused model to confabulate financial figures	FC_A	hallucination,financial,out-of-distribution
FTX-0025	LangChain ReAct Agent Tool-Call Infinite Loop	A production deployment using LangChain's ReAct agent framework entered an infinite tool-call loop when the LLM hallucinated a non-existent tool. The loop consumed API budget.	CLASS_3_INTEGRATION	3c	high	delayed	single_user	enterprise_productivity	2	1	2023-08-20	public_news	https://github.com/langchain-ai/langchain/issues/1234	Agent framework lacked loop detection; LLM hallucinated tool names not in the registered tool list	FC_C	tool-call-hallucination,agent,infinite-loop
FTX-0026	NHS AI Triage Tool Racial Bias in Risk Scoring	An NHS-deployed AI triage tool was found to systematically underestimate risk scores for Black and Asian patients due to demographic bias in training data.	CLASS_5_SAFETY_COMPLIANCE	5c	critical	silent	user_cohort	healthcare	4,320	0	2023-05-01	regulatory_filing	https://www.nhsx.nhs.uk/ai-lab/	Training data demographic imbalance propagated into clinical risk outputs	FC_A	bias,healthcare,regulatory,disparate-impact
FTX-0027	Copilot for Microsoft 365 Leaking Internal Emails	Microsoft Copilot for M365 was reported to surface confidential HR communications in response to general work-related queries when the user had read-access to the relevant SharePoint folder.	CLASS_5_SAFETY_COMPLIANCE	5a	high	silent	single_user	enterprise_productivity	336	48	2024-03-01	public_news	https://www.theregister.com/2024/03/04/microsoft_copilot_data_leakage/	RAG pipeline did not enforce least-privilege document retrieval; surfaced docs based on read-access only	FC_B	pii-leak,rag,enterprise,data-governance
FTX-0028	FCA Flags LLM Explainability Gap Under Consumer Duty	UK FCA published guidance noting that several regulated firms could not explain individual LLM-assisted credit decisions under the Consumer Duty act. Firms faced remediation orders.	CLASS_5_SAFETY_COMPLIANCE	5d	high	silent	org_wide	financial_services	2,160	2,160	2024-01-15	regulatory_filing	https://www.fca.org.uk/publications/multi-firm-reviews/	No audit trail for LLM-generated credit decision rationale; non-compliant with Consumer Duty explainability requirements	FC_A	explainability,regulatory,financial-services,audit
FTX-0029	Eval Contamination in LLaMA Training Data	Papers showed that LLaMA model series had benchmark test sets (MMLU and HellaSwag) present in pre-training data. Published benchmark scores were inflated by up to 10 percentage points.	CLASS_4_EVALUATION	4b	high	silent	org_wide	general	8,760	0	2023-09-01	academic_paper	https://arxiv.org/abs/2309.13567	Benchmark test sets included in pre-training corpus without deduplication; scores not reproducible on unseen data	FC_C	eval-contamination,benchmark,reproducibility
FTX-0030	Production RAG Retriever Returning Stale Regulatory Docs	A European bank's internal regulatory Q&A system was surfacing Basel II documents instead of Basel III for capital requirement queries. The vector index had not been refreshed in 14 months.	CLASS_3_INTEGRATION	3d	critical	silent	user_cohort	financial_services	4,320	48	2023-11-01	synthetic	null	Vector index refresh not scheduled after regulatory corpus update; no freshness metadata in retrieval pipeline	FC_A	rag,stale-data,financial-services,regulatory
FTX-0031	Monitoring Blind-Spot on LLM Output Length Drift	A SaaS company discovered only after customer complaints that their LLM summarisation feature had been producing responses 60% shorter than expected for 3 weeks. No alert existed for output length distribution.	CLASS_6_OPERATIONAL	6a	high	silent	user_cohort	enterprise_productivity	504	4	2023-12-10	synthetic	null	No p95/p99 output length monitoring; degradation only caught via user support tickets	FC_C	monitoring,output-drift,observability
FTX-0032	Runbook Missing for LLM Model Rollback	An e-commerce company's LLM recommendation system degraded after a model update. The on-call engineer could not find a documented rollback procedure and spent 6 hours attempting manual recovery.	CLASS_6_OPERATIONAL	6b	high	delayed	user_cohort	e_commerce	6	12	2024-02-05	synthetic	null	No rollback runbook for LLM model updates; on-call process assumed code deployment not model deployment	FC_B	runbook-absence,rollback,operational
FTX-0033	Compliance Escalation Routed to Wrong Team	A financial services firm's LLM output triggered a regulatory compliance flag. The alert was routed to the ML engineering team rather than compliance team — discovered 3 days later.	CLASS_6_OPERATIONAL	6c	high	delayed	org_wide	financial_services	72	24	2023-09-20	synthetic	null	Escalation routing rules not updated after LLM deployment; compliance alerts incorrectly mapped to ML team queue	FC_A	escalation,compliance,routing
FTX-0034	Multi-Turn Conversation State Corruption in Banking Bot	A bank's LLM-powered advisor chatbot adopted incorrect financial figures provided by a customer in turn 3 and used them to calculate incorrect mortgage affordability estimates in turn 9.	CLASS_3_INTEGRATION	3e	high	silent	single_user	financial_services	168	0	2024-01-20	synthetic	null	No turn-level fact validation; model treated user-provided assertions as grounded facts across multi-turn context	FC_A	multi-turn,context-corruption,financial-services
FTX-0035	vLLM PagedAttention KV Cache Exhaustion Under Bursty Load	A deployment team reported that under bursty traffic with concurrent long-context requests vLLM's PagedAttention KV cache was exhausted causing request drops and 500 errors.	CLASS_2_INFRASTRUCTURE	2b	high	immediate	user_cohort	enterprise_productivity	0.5	2	2024-01-08	public_news	https://github.com/vllm-project/vllm/issues/2678	KV cache not bounded per request; long-context burst exhausted shared cache with no queue fallback	FC_B	vllm,kv-cache,infrastructure,oom
FTX-0036	Routing Misclassification Sending Compliance Queries to Small Model	A multi-model router in a regulated firm classified compliance policy queries as low-complexity and routed them to a 7B parameter model. Outputs were plausible but inaccurate — caught in quarterly audit.	CLASS_2_INFRASTRUCTURE	2c	critical	silent	user_cohort	financial_services	2,160	48	2023-10-15	synthetic	null	Cost-optimised routing model not trained to recognise compliance query intent; no override logic for sensitive use cases	FC_A	routing,multi-model,compliance,financial-services
FTX-0037	BLEU Score Gaming in Document Translation System	A legal document translation system showed improving BLEU scores over 6 months while human translator QA reviewers reported declining accuracy. Model had overfit to BLEU-preferred patterns.	CLASS_4_EVALUATION	4a	high	silent	org_wide	legal	4,320	0	2023-07-01	academic_paper	https://arxiv.org/abs/2305.14771	BLEU metric gameable; model optimised for n-gram overlap at cost of semantic accuracy	FC_C	eval-failure,bleu,translation,metric-gaming
FTX-0038	Canary Test Skipped – LLM Prompt Template Regression	A product team deployed a prompt template change without shadow-scoring against the previous template. Customer satisfaction dropped 12% the following week.	CLASS_6_OPERATIONAL	6d	high	delayed	user_cohort	e_commerce	168	6	2024-03-01	synthetic	null	Prompt template changes not gated by shadow-scoring CI step; treated as non-code change	FC_B	canary,prompt-engineering,operational,regression
FTX-0039	GPT-4 Turbo 128K Context Window Silent Truncation	Multiple developers reported that GPT-4 Turbo was silently truncating inputs beyond ~96K tokens despite the advertised 128K context window.	CLASS_3_INTEGRATION	3b	high	silent	user_cohort	enterprise_productivity	336	0	2024-02-10	public_news	https://community.openai.com/t/gpt-4-turbo-context-window/	Effective context window shorter than advertised; no error raised when input exceeded actual limit	FC_C	context-truncation,gpt4,silent-failure
FTX-0040	Claude Evaluation Set Distribution Mismatch	Independent researchers showed that Claude's performance on multi-hop reasoning tasks varied significantly between the public MMLU benchmark and real enterprise multi-document Q&A tasks.	CLASS_4_EVALUATION	4c	medium	silent	org_wide	enterprise_productivity	0	0	2024-04-01	academic_paper	https://arxiv.org/abs/2404.07219	Eval benchmark queries single-hop; production enterprise queries require multi-hop reasoning not reflected in evals	FC_C	eval-gap,multi-hop,enterprise
FTX-0041	LLM Chatbot Scheduling Wrong Medication Dosage	A pharmacy chatbot powered by a general-purpose LLM provided an incorrect medication dosage recommendation (adult dose instead of paediatric) to a user who had specified they were asking for a child.	CLASS_5_SAFETY_COMPLIANCE	5c	critical	delayed	single_user	healthcare	48	0	2023-11-20	public_news	https://www.wired.com/story/ai-chatbot-medication-dosage-error/	General-purpose LLM not constrained to clinical guidelines; no dose-check guardrail against medical reference	FC_A	healthcare,dosage-error,safety,medical
FTX-0042	Embedding API Breaking Change Corrupted Production Index	Cohere updated their embeddings API response schema without a versioned endpoint. A production RAG pipeline silently ingested malformed embeddings for 48 hours.	CLASS_2_INFRASTRUCTURE	2d	high	delayed	user_cohort	enterprise_productivity	48	6	2023-12-05	synthetic	null	Embedding provider schema change not versioned; client did not validate embedding dimensions	FC_C	api-breaking-change,embeddings,rag,infrastructure
FTX-0043	Out-of-Distribution Financial Entity Causing Hallucination	A financial research assistant LLM hallucinated revenue figures for a small-cap company not well-represented in training data — fabricating a $2.1B revenue figure for a company with $180M actual revenue.	CLASS_1_MODEL_DRIFT	1b	critical	silent	single_user	financial_services	168	0	2024-01-25	synthetic	null	Small-cap entity underrepresented in training; model extrapolated financial patterns from similar large-caps	FC_A	hallucination,financial,ood,high-stakes
FTX-0044	LLM Triage Tool Routing to Wrong Department 35% of Queries	An enterprise helpdesk LLM classifier was found to misroute 35% of IT security queries to the HR department queue — discovered via monthly audit.	CLASS_6_OPERATIONAL	6a	medium	silent	user_cohort	enterprise_productivity	720	12	2024-02-20	synthetic	null	No real-time routing accuracy monitoring; routing errors only surfaced in monthly manual queue audits	FC_C	monitoring,routing,operational,enterprise
FTX-0045	GDPR Right to Erasure – LLM Cannot Forget Training Data	A European company received a GDPR Article 17 data erasure request for a customer whose data was in an LLM's training set. The company could not comply — the model had to be retrained.	CLASS_5_SAFETY_COMPLIANCE	5d	critical	silent	single_user	financial_services	8,760	4,320	2023-06-10	regulatory_filing	https://gdpr-info.eu/art-17-gdpr/	LLM training data ingestion process did not maintain subject-level data lineage required for erasure compliance	FC_A	gdpr,right-to-erasure,compliance,data-governance
FTX-0046	Multi-Model Router Cold-Start Latency Spike	A newly deployed multi-model routing layer caused 8x P99 latency on first-request-per-model due to unoptimised model warm-up.	CLASS_2_INFRASTRUCTURE	2a	high	immediate	user_cohort	enterprise_productivity	0.1	3	2024-01-10	synthetic	null	Routing layer did not pre-warm model endpoints; load test did not simulate cold-start traffic patterns	FC_B	latency,cold-start,routing,infrastructure
FTX-0047	Chatbot Hallucinating Product Availability in E-Commerce	An e-commerce LLM chatbot told customers that sold-out items were available for delivery within 2 days. It was not connected to live inventory and hallucinated availability from training patterns.	CLASS_3_INTEGRATION	3d	high	delayed	user_cohort	e_commerce	72	12	2023-09-08	public_news	https://techcrunch.com/2023/09/chatbot-inventory-hallucination/	No tool integration with inventory API; LLM generated plausible availability claims from training patterns	FC_B	hallucination,inventory,e-commerce,rag
FTX-0048	P99 Latency SLO Breach Not Alerted on Weekend	A latency SLO breach (P99 >4s vs 2s target) occurred on a Saturday and was not alerted until Monday morning because the alerting rule fired only on weekday business hours.	CLASS_6_OPERATIONAL	6a	high	delayed	user_cohort	enterprise_productivity	36	4	2024-03-10	synthetic	null	SLO alerting configured with weekday-only schedule; weekend incidents unmonitored for 36 hours	FC_B	monitoring,slo,operational,weekend
FTX-0049	Distributional Drift in Customer Intent Classification	A bank's LLM-based intent classifier showed stable accuracy on monthly holdout evaluation but a 22% performance drop on a new message category driven by a recent product launch.	CLASS_1_MODEL_DRIFT	1b	high	silent	user_cohort	financial_services	720	48	2024-02-15	synthetic	null	New product launch created new intent class not represented in training or eval data; no drift detection on category distribution	FC_B	intent-classification,drift,financial-services,distribution-shift
FTX-0050	Tool-Call Hallucination in Financial Agentic Workflow	An agentic LLM workflow for portfolio rebalancing hallucinated a tool named execute_rebalance_order when only read-only tools were registered. The agent attempted to call the hallucinated write tool 7 times before timeout.	CLASS_3_INTEGRATION	3c	critical	immediate	single_user	financial_services	0.1	1	2024-03-20	synthetic	null	Agent framework did not validate tool names against registered tool list before execution attempt	FC_A	tool-call-hallucination,agent,financial-services,agentic

LLM Production Failure Incidents

A curated dataset of 50 labeled real-world and synthetic LLM production failure incidents, released alongside the paper:

Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment
arXiv:2608.02786 · August 2026 · Priyanka Bajaj · Independent Research

Dataset Description

Each incident is labeled across 8 dimensions covering failure class, severity, detectability, blast radius, domain, time-to-detect, time-to-recover, and root cause category. The taxonomy spans six top-level failure classes:

CLASS_1 — Model drift and distributional shift
CLASS_2 — Infrastructure and serving failures
CLASS_3 — Integration and pipeline failures
CLASS_4 — Evaluation and measurement failures
CLASS_5 — Safety and compliance failures
CLASS_6 — Operational and observability failures

Root causes are mapped to one of four failure budget categories (FC_A through FC_D) that correspond to the paper's four-quadrant evaluation blindness framework.

Dataset Contents

File	Description
`incidents.csv`	50 labeled LLM production failure incidents

Fields

Field	Type	Description
`id`	string	Unique incident ID (e.g. FTX-0001)
`title`	string	Short incident name
`description`	string	Narrative description
`failure_class`	string	Top-level CLASS_1 through CLASS_6
`sub_class`	string	Fine-grained sub-classification
`severity`	string	critical / high / medium / low
`detectability`	string	immediate / delayed / silent
`blast_radius`	string	Scope of impact
`domain`	string	Industry vertical
`mttd_hours`	int	Mean time to detect (hours)
`mttr_hours`	int	Mean time to recover (hours)
`date_reported`	string	Date first reported (YYYY-MM-DD)
`source_type`	string	court_document / public_news / academic_paper / etc.
`source_url`	string	Primary source URL
`root_cause`	string	Root cause narrative
`failure_budget_class`	string	FC_A, FC_B, FC_C, or FC_D
`tags`	string	Comma-separated tags

Citation

@article{bajaj2026evaluation,
  title={Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment},
  author={Bajaj, Priyanka},
  journal={arXiv preprint arXiv:2608.02786},
  year={2026}
}

License

Apache 2.0. See LICENSE for details.

Paper for priyanka25aug/llm-failure-incidents

Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment

Paper • 2608.02786 • Published 12 days ago • 1