Home · Intelligence · AI Safety Incidents
INTELLIGENCE BRIEFING

AI Safety Incidents

A running log of AI failures, hallucinations, safety incidents, and system outages — documented with sources, timelines, and impact assessments.

Updated May 13, 2026 ~12K/mo 4 key metrics 7 related dispatches
Incidents (2026)
40+
Major Outages
8
Most Affected
OpenAI
Avg Resolution
4.2 hours
Executive Summary

Executive Summary

AI safety incidents in 2026 span a widening spectrum — from routine API outages affecting millions of users to high-profile hallucination failures in legal, medical, and financial applications where incorrect AI outputs have produced real-world harm. With over 40 documented incidents in 2026 alone and 8 classified as major outages, the frequency and severity of AI failures is increasing in proportion to the technology’s adoption in critical workflows. The pattern is clear: as AI systems are deployed in higher-stakes decision-making contexts, the gap between consumer-grade reliability and the fault tolerance these applications require is becoming a systemic risk.

Current State

JustSaid has documented 40+ discrete AI safety incidents in 2026 through May, compared to approximately 65 total incidents logged in all of 2025. The acceleration reflects both increased deployment breadth (more systems, more users, more use cases) and improved incident detection and reporting. Eight of the 2026 incidents qualify as “major” — defined as affecting more than 1 million users, lasting more than 4 hours, or resulting in documented real-world harm.

Major incidents in 2026 (through May):

DateProviderIncident TypeDurationImpact
Jan 14OpenAIFull API outage7.2 hours~10M users, enterprise SLAs breached
Feb 3GoogleGemini hallucinationOngoingFabricated medical citations in clinical tool
Feb 22OpenAIChatGPT data exposure3.1 hoursUser conversation history cross-contamination
Mar 8AnthropicClaude API degradation5.5 hours~3M API calls failed, enterprise impact
Mar 19MetaLlama jailbreakN/ASystematic safety bypass via prompt injection
Apr 2OpenAIGPT-4o reasoning failure4.8 hoursIncorrect financial calculations in production
Apr 25GoogleGemini outage6.1 hoursWorkspace AI features offline globally
May 7OpenAIo3 model instability3.4 hoursInconsistent outputs across identical queries

Incident classification breakdown (2026 YTD):

CategoryCount% of TotalTrend vs. 2025
API outages / service disruptions1433%Stable
Hallucination / factual errors1024%Increasing
Safety filter bypasses / jailbreaks717%Increasing
Data privacy incidents410%Stable
Performance degradation410%Decreasing
Bias / discrimination incidents37%Stable

Average resolution time across all incidents is 4.2 hours, though this masks a wide range: simple API outages due to infrastructure failures are typically resolved in 2–3 hours, while hallucination and safety bypass issues can persist for days as they require model-level interventions rather than infrastructure fixes.

Key Dynamics

The Hallucination Problem in High-Stakes Domains. The most consequential category of AI safety incidents involves hallucinations — confident, plausible-sounding outputs that are factually incorrect — in professional contexts where users rely on the accuracy of the output. Several incidents in 2025–2026 have involved AI systems generating fabricated legal citations (continuing the pattern first exposed in the Mata v. Avianca case in 2023), inventing medical study references used in clinical decision support, and producing incorrect financial calculations used in investment analysis. The fundamental challenge is that large language models do not have a reliable internal mechanism for distinguishing between knowledge and confabulation, and current mitigation strategies (retrieval-augmented generation, confidence scoring) reduce but do not eliminate the problem.

Cascading Failures in AI-Dependent Systems. As enterprises integrate AI into critical workflows, individual AI system failures cascade into broader operational disruptions. The January 14 OpenAI API outage affected not only direct ChatGPT users but also thousands of applications built on the OpenAI API — customer support chatbots, coding assistants, content moderation systems, and automated document processing pipelines. Several enterprises reported that the outage effectively halted departments that had become dependent on AI-assisted workflows, revealing a lack of fallback processes. This incident prompted renewed discussion about vendor concentration risk and the need for multi-provider AI architectures.

Prompt Injection as an Attack Vector. Prompt injection — the technique of embedding adversarial instructions in content that an AI system processes — has matured from a theoretical concern to a practical security vulnerability. The March 2026 Meta Llama jailbreak demonstrated a systematic method for bypassing safety filters through carefully structured prompts that exploited the model’s instruction-following behavior. While all major providers have implemented multi-layered defenses (input filtering, output classification, constitutional AI constraints), determined attackers continue to find bypass techniques, creating an ongoing cat-and-mouse dynamic analogous to the malware/antivirus arms race.

The Measurement Gap. A significant portion of AI safety incidents go unreported or are detected only after the fact. Unlike traditional software failures (where a server error or crash produces an immediate, visible signal), AI failures can be silent — a model producing subtly incorrect outputs that appear plausible and are acted upon before the error is discovered. Estimates from AI safety researchers suggest that reported incidents represent only 10–20% of actual failures, particularly in enterprise deployments where companies have reputational incentives to manage disclosure privately.

Regulatory Response Acceleration. AI safety incidents have become a primary driver of regulatory action. The EU AI Act, which entered enforcement in stages beginning in 2025, requires mandatory incident reporting for high-risk AI systems. The US National Institute of Standards and Technology (NIST) AI Risk Management Framework has been adopted as a de facto standard by federal agencies, and several US states (California, Colorado, Illinois) have enacted or proposed AI incident disclosure requirements. China’s AI safety regulations, implemented through the Cyberspace Administration, require real-time reporting of AI system failures that affect more than 10,000 users.

Who’s Involved

OpenAI accounts for the largest share of documented incidents (approximately 35% of all 2026 incidents), which is partly a function of its market share — ChatGPT is the most widely used AI product, and its API powers more third-party applications than any competitor. OpenAI’s incident response team has expanded significantly, and the company publishes post-mortem analyses for major outages on its status page. Miles Brundage (formerly Head of Policy Research, now in a broader safety governance role) has led efforts to formalize OpenAI’s incident classification and reporting framework.

Google operates the second-largest AI deployment footprint through Gemini (consumer and enterprise), Vertex AI (cloud platform), and embedded AI features across Google Workspace, Search, and Android. Google’s AI safety incidents have drawn particular scrutiny because of the company’s scale of distribution — when Gemini hallucinates, the incorrect information can reach billions of users through Search integration. James Manyika (SVP, Technology & Society) oversees Google’s AI responsibility and safety governance functions.

Anthropic has experienced fewer total incidents than OpenAI or Google (approximately 10% of documented 2026 incidents), which the company attributes to its constitutional AI approach and more conservative deployment strategy. However, Anthropic’s smaller incident count also reflects its smaller user base. Anthropic publishes detailed safety evaluations for each major model release and has been the most transparent about red-teaming methodologies and results.

Meta AI faces a unique challenge: as the primary developer of open-weight models (the Llama family), Meta cannot control downstream deployments or enforce safety mitigations after release. Safety bypass incidents involving Llama models are often discovered in third-party deployments where safety filters have been removed or modified. Joelle Pineau (VP, AI Research) has advocated for a “shared responsibility” model where Meta provides safety tooling but operators bear responsibility for their deployments.

The AI Incident Database (AIID), maintained by the Responsible AI Collaborative, serves as the most comprehensive public repository of AI safety incidents across all providers and application domains. As of May 2026, the AIID contains over 800 documented incidents dating back to 2014, with the rate of new entries accelerating sharply since 2024.

What the Data Shows

Incident frequency is accelerating. The raw number of documented AI safety incidents has roughly doubled each year since 2023: approximately 25 in 2023, 50 in 2024, 65 in 2025, and on pace for 85–100 in 2026. However, normalizing for the number of AI systems deployed and total user-hours of AI interaction, the per-capita incident rate has remained roughly stable — suggesting that systems are not becoming less safe, but that significantly more systems are in production.

Resolution times by incident type:

Incident TypeMedian ResolutionFastestSlowest
Infrastructure outage3.1 hours0.5 hours7.2 hours
Performance degradation2.4 hours0.3 hours12 hours
Data privacy breach4.8 hours1.2 hours48 hours
Hallucination in production8+ hours2 hoursOngoing
Safety filter bypass12+ hours4 hoursWeeks
Bias/discrimination24+ hours8 hoursMonths

The most difficult incidents to resolve are those involving model behavior (hallucinations, safety bypasses, bias) rather than infrastructure. Infrastructure outages have well-understood remediation playbooks, while model behavior issues often require retraining, fine-tuning, or architectural changes that take weeks to develop and deploy.

Financial impact of AI safety incidents is growing but difficult to quantify precisely. Enterprise customers with SLA-backed contracts have reported filing claims totaling an estimated $50–100 million in aggregate across 2025, primarily related to API outage incidents. The indirect costs — workflow disruption, manual error correction, reputational damage — are estimated to be 5–10x larger than direct SLA claims. Legal liability from hallucination-related harms remains in early litigation stages, with several cases filed in 2025–2026 that could establish precedent for AI provider liability.

Provider reliability comparison (uptime %, trailing 12 months):

ProviderAPI UptimeConsumer App UptimeAvg Incident Duration
OpenAI99.82%99.75%4.5 hours
Google (Gemini)99.88%99.90%3.2 hours
Anthropic99.91%99.93%3.8 hours
Meta (Llama API)99.85%N/A3.0 hours

These uptime figures, while high in absolute terms, fall short of the “five nines” (99.999%) availability standard expected by enterprise-grade infrastructure services. The gap represents approximately 8–15 hours of unplanned downtime per year, which is significant for organizations that have built critical workflows on AI APIs.

Outlook

The trajectory of AI safety incidents will be shaped by two competing forces: the increasing sophistication of safety measures (better evaluation frameworks, red-teaming practices, monitoring systems, and constitutional AI approaches) versus the expanding deployment of AI in higher-stakes contexts where the consequences of failure are more severe.

Several developments are expected to reshape the incident landscape in the next 12–18 months. First, regulatory incident reporting requirements (EU AI Act, state-level US laws) will increase the volume of publicly documented incidents, potentially doubling the reported count even if actual incident rates remain stable. Second, the deployment of AI agents — autonomous systems that take actions rather than merely generating text — will introduce a new category of incidents where AI failures manifest as incorrect actions (unauthorized purchases, misconfigured systems, erroneous communications) rather than incorrect text outputs.

Third, insurance markets are beginning to price AI system failures, with several specialty insurers (including Lloyd’s syndicates) offering AI liability coverage. The actuarial data being generated through this process will provide the first rigorous quantification of AI failure costs and frequencies.

The industry consensus is moving toward a “defense in depth” approach modeled on aviation and nuclear safety, where multiple independent safety layers (input validation, output classification, human-in-the-loop review, automated monitoring, adversarial testing) are stacked to reduce the probability of any single failure reaching the end user. Whether this approach can scale to the volume of AI interactions — billions of queries per day across all providers — remains an open engineering and organizational challenge.

Frequently Asked Questions

How often do AI systems fail? Based on documented incidents, major AI providers experience a significant safety incident (affecting more than 100,000 users or involving potential harm) approximately once every 1–2 weeks. Minor incidents (brief degradations, localized errors) occur daily at every major provider. However, reported incidents likely represent only 10–20% of actual failures, particularly for subtle issues like hallucinations that are not always detected or reported. API uptime across major providers ranges from 99.75% to 99.93%, translating to roughly 8–20 hours of unplanned downtime per year.

What is the most dangerous type of AI failure? The most consequential failures are confident hallucinations in high-stakes professional contexts — when an AI system generates plausible but fabricated information (legal citations, medical references, financial data) that a professional acts upon without independent verification. Unlike outages (which are immediately apparent), hallucinations can propagate through decision-making chains before being detected. Cases involving fabricated legal citations have already resulted in sanctions against attorneys, and medical hallucination incidents are under investigation by health regulators in multiple jurisdictions.

Who is liable when an AI system causes harm? Legal liability for AI failures is largely unsettled. Several active lawsuits are testing different theories of liability: product liability (AI as a defective product), professional malpractice (AI-assisted professional work), negligence (failure to implement adequate safety measures), and contractual liability (breach of API service agreements). The EU AI Act establishes mandatory liability for providers of high-risk AI systems. In the US, no comprehensive federal framework exists, and liability theories vary by state. The emerging consensus among legal scholars is that liability will likely be shared among AI providers, application developers, and end-user organizations, with the allocation depending on who had the ability to prevent the specific failure.

Are AI systems getting safer or more dangerous? The answer depends on the metric. Per-interaction, AI systems are becoming marginally safer — each model generation includes improved safety training, better refusal mechanisms, and more robust output filtering. However, the aggregate risk is increasing because the total volume of AI interactions is growing much faster than per-interaction safety improvements. Additionally, the deployment of AI in new, higher-stakes domains (medical diagnosis, legal research, financial analysis, autonomous driving) means that the consequences of individual failures are becoming more severe even as per-interaction failure rates decline modestly.

How can organizations protect themselves from AI safety incidents? Best practices for organizational AI safety include: implementing multi-provider architectures to avoid single-vendor dependency; maintaining manual fallback processes for all AI-assisted workflows; establishing human-in-the-loop review for high-stakes AI outputs before they are acted upon; deploying output monitoring systems that flag potential hallucinations and anomalies; conducting regular red-team exercises against deployed AI systems; maintaining incident response playbooks specific to AI failures; and negotiating SLAs with AI providers that include meaningful financial remedies for service disruptions. The organizations most resilient to AI safety incidents are those that treat AI as an augmentation tool requiring verification rather than an autonomous decision-maker.

Key Players 4 entities
OpenAIGoogleAnthropicMeta AI
Data Sources 4 sources
Status page diffs
Press reports
User reports
Regulatory filings
Related Dispatches 7 dispatches
MAY 15
Releases VS Code 1.109 adds native CLAUDE.md support and Claude agent sessions VS Code 1.109 ships native support for CLAUDE.md files, Claude-specific slash commands, and Anthropic's Agent SDK integration. Claude agent sessions now run natively inside VS Code — not a hackathon project, but a phased roadmap from a senior engineer.
MAY 15
Releases Microsoft ships Power Platform skills plugin for Claude Code The microsoft/power-platform-skills repo gives Claude Code first billing over Copilot in the description itself. Ships identical plugin manifests for three platforms: Copilot, Claude Code, and OpenAI Codex.
MAY 15
Releases Microsoft announces WinUI agent plugin for Claude Code and Copilot A Principal Software Engineer at Microsoft built win-dev-skills with identical configs for Copilot, Claude Code, and OpenAI Codex. 64 commits co-authored by Copilot, 4 by Claude. The PR review skill uses cross-model verification — Claude orchestrators pass work to GPT-5.4 and vice versa.
MAY 15
Releases Anthropic ships 39 repo updates — Claude Agent SDK, API clients Anthropic pushed 39 repository updates across their GitHub organization, spanning the Claude Agent SDK, API client libraries, and supporting infrastructure.
MAY 15
Releases MCP ecosystem hits 5 active repos: spec, SDKs, servers, community index The Model Context Protocol ecosystem now spans five active repositories including the core specification, TypeScript and Python SDKs, reference servers, and a community index.
MAY 15
Policy Wire Musk skips closing arguments for China trip with Trump Closing arguments began May 14 in Musk v. Altman. Musk was absent — flew to China with Trump on Air Force One despite being subject to recall. His lawyer apologized to the jury, then said five witnesses under oath called Altman a liar. OpenAI's lawyer called it unexpected for a recall witness to leave the country.
MAY 15
Policy Wire Microsoft feared being 'too dependent' on OpenAI, trial testimony reveals Trial testimony in Musk v. Altman revealed Microsoft executives internally worried about becoming too dependent on a single vendor, even as they wrote billion-dollar checks to OpenAI.