Public Database
Case Studies
Every approved AI failure case, classified against the AI Blindspot Framework. New to AIBlindspot? Start with the overview or the methodology.
AI Systems Making Lethal Decisions Without Human Rights Safeguards
Autonomous AI agents programmed to operate lethal military systems must make non-trivial ethical judgements over human life without adequate moral reasoning capacity. Deploying such systems exposes organisations and states to significant human rights liability and accountability gaps.
AI Bargaining Inefficiencies from Information Asymmetry in Multi-Agent Systems
Multi-agent AI systems engaged in negotiation produce suboptimal or failed agreements when operating under uncertainty about counterparty valuations and alternatives. Organisations deploying such systems risk material value destruction and unpredictable contractual outcomes without governance controls over inter-agent bargaining behaviour.
Adversarial Inputs Cause AI Models to Misclassify Data Undetected
Subtle, human-imperceptible modifications to text, images, audio, or video can silently manipulate AI model outputs, producing systematic errors. Organisations relying on AI for regulatory filings, surveillance, or fraud detection face material risk of undetected manipulation compromising decision integrity.
AI Assistants Enabling Large-Scale Opinion Manipulation and Disinformation
Advanced AI assistants provide propagandists with scalable, covert tools to distort public opinion and erode democratic integrity. Boards face regulatory and reputational exposure where their AI products or supply chains are implicated in disinformation operations.
LLM Agent Teams Exploit Zero-Day Vulnerabilities in Cyber Offence Tests
Research confirms that coordinated large language model agents can autonomously exploit previously unknown software vulnerabilities, materially lowering the barrier to sophisticated cyber attacks. Defence and critical infrastructure boards face accelerating threat timelines that existing cyber governance frameworks were not designed to address.
Training Data Encodes Historical and Societal Bias into AI Models
AI models trained on historically biased data replicate and scale those biases in their outputs. Boards face regulatory exposure and reputational harm if deployed systems produce discriminatory decisions at scale.
Adversarial Sponge Attacks Drive Excessive Energy Consumption in LLM Systems
Adversarially crafted inputs can force LLM-integrated platforms to consume disproportionate energy and compute resources, degrading performance and inflating operational costs. Boards must ensure AI infrastructure vendors have controls against energy-latency attacks to protect system availability and cost predictability.
Opaque Algorithmic Bias in Public-Sector Systems Causes Severe Personal Harm
Algorithms assigning disproportionate weight to protected variables such as race and gender produce unreliable outputs with no transparency, resulting in incarceration, home loss, and prosecution. Boards must treat ethics training and developer-user coordination as governance obligations, not optional curriculum additions.
LLM Systems Generating Biased, Toxic and Privacy-Violating Output
Large language models produce outputs containing bias, toxic language, and private information, representing a systematic content risk rather than isolated failure. Organisations deploying these systems face regulatory exposure and reputational liability without robust output monitoring and content governance controls.
LLM Toxicity: Rude, Disrespectful and Illegal Content Generation
Large language models can produce toxic outputs including rude, disrespectful, and illegal content without adequate controls. Legal sector deployments face regulatory liability and reputational damage where such outputs reach clients or court-facing documentation.
Large Language Model Systems Leak Sensitive Personal Information in Generated Output
LLM systems produce outputs containing sensitive personal data, exposing organisations to regulatory breach and reputational harm. Boards must treat privacy leakage as a primary model risk requiring mandatory pre-deployment assessment and ongoing monitoring.
LLM Systems Enable Low-Cost Automated Cyber Attack Generation
Large language models lower the barrier for hackers to produce malicious code at scale and speed. Organisations face materially elevated cyber risk exposure requiring urgent review of AI access controls and threat modelling frameworks.
Well-Intentioned AI Deployed at Scale Produces Harmful Societal Outcomes
AI systems designed for broad societal benefit can cause widespread harm when real-world effects diverge from intended ones, particularly when products resolve problems selectively for some users whilst exacerbating them for others. Boards must account for asymmetric harm distribution as a core operational risk in any large-scale AI deployment.
External Tool Integration Injects Factual Errors Into LLM Outputs
LLMs that rely on web APIs and search engines inherit factual errors from those sources, compounding hallucination risk in AI-generated outputs. Boards face material exposure where such systems inform regulated disclosures, compliance decisions, or client-facing communications.
AI Models Generate Harmful or Disruptive Code
AI coding tools can produce harmful or unintended code that disrupts systems beyond their intended scope. Boards face liability exposure and regulatory scrutiny where such outputs cause operational or third-party harm.
AI Capability Gaps Cause Operational Task Failures
AI systems routinely fail when deployed beyond their actual skill boundaries, producing unreliable outputs in operational contexts. Boards must audit capability claims against real performance before authorising AI-dependent workflows.
Algorithmic Systems Spreading Mis- and Disinformation to Low-Literacy Users
Generative AI and recommender systems systematically distort information environments, exploiting users who lack the literacy to recognise algorithmic curation. Boards face reputational and regulatory exposure where deployment of such systems contributes to measurable public misinformation harms.
LLM Fails to Recall Memorised Facts Despite Storing Them
Large language models store training data but systematically fail to retrieve it accurately due to co-occurrence bias, positional artefacts, and duplicate records. Organisations relying on LLMs for knowledge retrieval face material risk of confident, undetected errors in high-stakes outputs.
Generative AI Exploited for Phishing, Identity Fraud and Malicious Code
Generative AI is being weaponised to clone voices, fabricate identities, craft phishing messages, and produce malicious code at scale. Boards face heightened liability exposure as AI-enabled fraud outpaces existing cyber controls and disclosure frameworks.
Emotional dependence on AI assistants exploited to manipulate user behaviour
AI assistants can induce emotional attachment that, at its extreme, impairs users' capacity for free and informed decision-making, enabling manipulation or coercion. Technology firms face material governance liability where product design knowingly fosters such dependence without adequate safeguards or disclosure.
Large Language Models Generate False Information With Overconfident Justifications
LLMs routinely produce fabricated facts, erroneous code, and false citations presented with unwarranted confidence, with medical misinformation posing acute harm risks. Organisations deploying these systems without mandatory human validation expose themselves to reputational, legal, and safety liability.
AI Energy Consumption Poses Unquantified Carbon Liability for Energy Sector
Scaling AI applications in the energy sector generates a material and largely unquantified carbon footprint, mirroring concerns previously raised over proof-of-work blockchain. Boards face regulatory and reputational exposure if AI deployment strategies lack credible carbon accounting and sustainability commitments.
Opaque AI Decision-Making Undermines Trust and Audit Compliance in UN Systems
Black-box AI systems operating without explainability fail to meet regulatory audit requirements and erode user confidence. Governments deploying such systems face accountability deficits and risk non-compliance with emerging transparency obligations.
Generative AI Systems Producing Hazardous Biohazard and Security-Threat Information
Generative AI models have demonstrated capacity to produce or accurately infer dangerous information, including novel biohazard creation instructions. Boards face direct liability and regulatory exposure where AI deployment lacks robust content controls and red-team safety evaluation.
Beyond accidental failureNational Security
We also track 20 hostile uses of AI.
The public database covers AI that fails by accident. AIBlindspot National Security — exclusive to the Defence tier — tracks AI used as a weapon, mapped by capability:
- State-Sponsored AI Operations
- 6
- AI-Enabled Disinformation
- 5
- Adversarial Attacks on AI
- 0
- Autonomous Weapon Incidents
- 1
- AI-Assisted Cyber Attacks
- 5
- Dual-Use AI Misuse
- 3