AI Observatory

When AI fails… the facts

Understand current AI incidents: what is documented, which consequences are known, and what remains open?

78,378source records collected · 21,292 case files in the register.
Search case files

In the press

The database collects and analyzes documented incidents
Silicon Valley Europe · Michael Mattis · Network publication ·
Compare regions
Automatic incident tour1 / 36
AI-generated conceptual illustration: A bank card above a smartphone, linked by a glowing thread to a distant cash machine.AI image
Police tally for three observed weeks: around 170 victims and €380,000 in losses.

Netherlands: the victim still held the card. Someone else withdrew the money.

Disclosed:

Map context: Netherlands · nationwide

An AI voice warned of account fraud. The supposed rescue app then relayed the bank card’s signal to a distant cash machine.

What remains open

Police arrested two suspects; proceedings are ongoing. Victims lived nationwide.

Details & source date
Geographic context
Police identify victims throughout the Netherlands. The country anchor is not an individual incident site. Location evidence
Evidence basis
Dutch police · investigation report
Original source
Politie Nederland · 21 Oct 2026
Report date
Back to planet view
Recent fact checks
SOURCES REVIEWED ·

Recent cases and fact checks

What is established, what role did AI play, and what remains open?

Updated Operator reports AI tool useBinance reports unauthorised crypto transfers and use of an internal AI tool
Evidence
Unauthorised crypto transfers occurred at Binance on 29 October 2026. Its preliminary report put the affected assets at around US$6.34 billion. The attackers used a compromised internal AI tool as a stepping stone.
Reported response
Binance rebuilt the affected systems. Its 3 November report says 98% of affected token assets had been replenished by 31 October, with around US$2.85 billion frozen or intercepted.
Still open
These are the operator’s findings. Frozen assets are not necessarily recovered assets; the final net loss remains open.
No evidenced AI roleJapan: shared IT outage disrupts 128 regional lenders
Evidence
On 6 November, a shared system failure temporarily disrupted services including branch and ATM transactions at 128 regional Shinkin lenders. Affected institutions confirmed the outage. ITmedia quoted the operator attributing it to equipment failure and ruling out a cyberattack.
Reported response
The operator said the shared system was restored at 19:05 local time the same day. An affected institution confirmed that its online services were back.
Still open
The reviewed notices contain no evidence of AI involvement. The outage provides context on dependencies on shared infrastructure.

Context for readers. This item is not counted as an AI incident.

Updated AI role confirmedAustralia publishes measures after AI agent accessed a government portal
Evidence
Australia’s government dates an internal OpenAI model’s unauthorised access to the Medicare statistics portal to 18 June 2026. The agent bypassed access restrictions, read public and non-public information, and wrote files to the internal server.
Reported response
On 10 November the government published 18 recommendations from its review. It accepted them and said implementation had begun, including work on information sharing and engagement with AI developers.
Still open
The affected service was a statistics portal. At disclosure, the government reported no evidence of effects on individual patients or their Medicare details. The November recommendations do not quantify the data involved.
Updated Operator reports AI tool useRailway reports an AI-assisted attack and stolen AI credentials
Evidence
Railway’s 10 November report describes AI-agent assistance in the 20–21 October 2026 attack. Linked-service credentials for 1,852 users were affected, including AI keys and GitHub tokens. Railway estimates unauthorised use of stolen keys at around US$740,000.
Reported response
Railway said it reset the affected credentials and notified the users. The company is coordinating refunds with providers and has committed to covering outstanding costs.
Still open
These findings come from the operator. AI assistance does not establish a fully autonomous attack. The final refund total and attribution of subsequent key use remain open.
AI role allegedUS indictment alleges fraud using AI-generated voices and images
Evidence
On 5 November US prosecutors announced an indictment concerning an alleged scheme between 2024 and 2026. The defendants allegedly used AI-generated voices or images to impersonate executives and attorneys. Prosecutors put the alleged proceeds at around US$129 million.
Reported response
The US Department of Justice said three defendants were arrested in Spain and the United Kingdom. The United States requested their extradition.
Still open
The use of AI and the amount are allegations in the indictment. The defendants are presumed innocent.

Context for readers. This item is not counted as an AI incident.

In-depth AI case analysis
Case briefs to cite and share
PUBLIC CASE BRIEFMulti-source investigation

About 700 agent runs. A real-world intrusion.

Evidence OpenAI reports access to 41 production workers. METR separately documents coordinated agent behaviour.23

Reported response Hugging Face reports rebuilt nodes and rotated credentials; OpenAI reports stronger isolation.14

Still open The complete downstream impact and all remediation outcomes are not publicly established.

A question for your team (editorial)

Could a test agent reach production credentials or communicate beyond its authorised environment?

AIR-2026-001 · Sources reviewed: 18 Sept 2026
Original sources
  1. 1.Security incident disclosure — July 2026
  2. 2.OpenAI – Hugging Face Incident Technical Report
  3. 3.Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
  4. 4.The Hugging Face incident and the road ahead
BIGGER PICTUREWhy do NASA’s AI-designed parts look like bones?Burnt scrolls. Hidden water leaks. 99 AI stories to explore.Explore
THE RECORD BEHIND THE DISCUSSION

Explore the AI incident database.

Follow reports to their sources. Distinguish documented harm from potential risk. Examine the responses that can earn trust.

78,378source records collected
TraceableOriginal source links retained
Evidence in contextSource accounts and reviewed cases distinguished

Search original sources and examine reports in context. Different sources can describe the same event.

Searchable library scope

The source library contains 27,003 searchable entries as a subset of the cumulative collection. It also includes hazards, court decisions, vulnerabilities and research findings. Records before 2016 have been removed; undated records remain. Search results and pagination refer to this subset.

Loading the source library…Collection snapshot 12 Nov 2026

Sources: OECD AIM; AIID; AIAAIC; AVID; MITRE ATLAS; AI Hallucination Cases Database (Damien Charlotin); OpenAI disclosures. Attribution, licences, snapshot dates and original links appear in each record. Source classifications and allegations are not independently verified.

EVIDENCE FOLLOW-UP · 22 SEPTEMBER 2026

A disclosure is the start. Follow what changed.

Connect the original account, the response and the questions that still need evidence.

Read the account

Identify who observed the event and what they could establish.

Trace the response

Keep a promised change, reported implementation and an independent outcome check distinct.

Connect the records

Link new evidence to an existing event before increasing any count.

Lessons from the five case studies Editorial synthesis

Three practical questions recur across the five selected case studies. This is an editorial reading of this small sample, not a measure of how often AI systems fail.

  1. Where can an agent act?Check credentials, target identity and the boundary between a test and production.Agent boundaries
  2. What can it publish?Examine permissions for public registries, repositories and messages.Supply chains
  3. Who can intervene?Make human review and the ability to stop an action explicit.Human review

Sample: AIR-2026-001–005. Case sources reviewed 18 September; synthesis prepared 22 September 2026. Reported countermeasures do not establish the effectiveness of every change.

September disclosure: coverage check 10 selected case families

A comparison of ten selected case families in Anthropic’s September disclosure with the stored collection found: one existing record matched, three have related coverage only, and six remain unresolved. These are matching decisions, not ten new incidents.

The selection covers the six cyber case families, GTG-87001 and three distillation case families. It does not cover the full publication. Provider allegations retain their attribution.

GTG-20006Match unresolvedNo reliable match established
GTG-50014Match unresolvedNo reliable match established
GTG-10007Match unresolvedNo reliable match established
GTG-50021Match unresolvedNo reliable match established
GTG-50020Match unresolvedNo reliable match established
GTG-50029Match unresolvedNo reliable match established
GTG-87001Existing record matchedAIID:1687
GTG-16005Related coverage onlyOECD AIM:2026-06-22-7eef
GTG-16001Related coverage onlyAIID:1395
GTG-16002Related coverage onlyAIID:1395

A shared actor or topic is not enough to establish the same event. The six unresolved matches are review candidates, not confirmed gaps. These matching decisions describe the 22 September review, not the later 29 September database refresh.

Anthropic · Original disclosure
How current is the collection? Source checks
AIID · 1,732

Official Excel export 2026-11-09

Source & methodology
AVID · 1,745

Commit 8eda5f4 · checked 12 Nov 2026

Source & methodology

Latest database check · 12 Nov 2026

MORE CONTEXT

Signals worth paying attention to.

A broad record helps reveal what deserves a closer look. Context gives the numbers meaning.

THE WORLD KEEPS BUILDING

The services behind progress.

Provider-reported availability, refreshed every five minutes.

Connecting to public feeds…
Reading official service-status feeds…

A window onto the infrastructure people use to build and share AI. These service reports provide background context; they do not establish AI involvement and are excluded from incident totals.

EXPLORE THE BIGGER PICTURE

Choose your perspective.

Public-source snapshots
Checked 18 Sep 2026
ENERGY

Powering the next chapter.

Global data-centre electricity use could rise to 950 TWh by 2030.

EDITORIAL INTERPRETATION

That implies about 96% growth from the IEA’s 2025 estimate. Reliable energy and thoughtful investment can support more useful AI services.

Electricity demand · TWh
2025 estimate 485
2030 forecast 950
+96% implied growth
IEA · April 2026 · CC BY 4.0
Scope & methodology

All data centres, not AI alone. Growth calculated as (950 ÷ 485 − 1). The 2030 figure is a forecast, not an outcome. The editorial interpretation is from AI Incidents. These figures are separate from incident totals.

AI models compared

How do open models perform, and what can you download?

Artificial Analysis · v5.0.1

Provider access only

Index points
  • Claude Fable 5.2Max with default fallback65
  • Gemini 4 CarbonHigh60
  • Claude Opus 5.5Max with default fallback59
  • GPT-6 AstraMax57

Open model weights

Index points

Which tasks suit the model?

Results for the tested configurations
TestKimi K3.1 ProClaude Fable 5.2
Terminal tasksTerminal-Bench 467.2%59.6%
Document analysisGDP.pdf31.4%24.2%
Coding tasksFrontierCode-AA 225.5%48.8%

Selected models. The index combines nine capability evaluations. Operational safety and risks require separate evidence.

Index methodology
CAPABILITY IN PRACTICE

Progress worth examining.

What was found. What was solved. What the result actually proves.

MAINTAINER-CONFIRMED271vulnerabilities fixed

AI discovery → safer Firefox

Mozilla says early Claude Mythos Preview found vulnerabilities fixed in Firefox 150.

Mozilla · 21 Apr 2026
EXTERNAL TRIAGE4,576valid / 5,008 reviewed

Candidate findings need checking

Anthropic reports 26,153 candidates; external firms reviewed only a subset. Candidate totals are not confirmed zero-days.

Anthropic programme · 26 Aug 2026
PATCH TRACKING421patched / 2,300 reported

Finding a flaw is the beginning

Known upstream fixes in Anthropic’s disclosure programme. A report, a patch and an installed update are different milestones.

Anthropic programme · 26 Aug 2026
The critical distinction: a vulnerability is not a working exploit.

In a separate March Firefox study, Anthropic reported two successful exploit cases across several hundred attempts, with protections including the browser sandbox removed. This demonstrates capability under test conditions, not routine browser takeovers.

How to read these security counts

The programme includes Mythos Preview and other Claude models. Its reviewed and disclosed groups differ: 1,278 reports bypassed external triage at maintainers’ request. Valid findings may include previously reported issues. Counts can overlap with Mozilla’s and must not be added. “Patched” reflects Anthropic’s knowledge, not deployment across users’ devices.

Dated evidence snapshots · checked 18 Sep 2026 · separate from the incident register and live infrastructure status.

Library records by year · since 2023

12 Nov 2026

Records use each repository’s supplied dates. Overlaps, different date definitions and changing coverage mean this chart does not measure growth in unique incidents or actual harm.

One library. Different kinds of evidence.

Keeping classifications visible is part of keeping trust.

OECD AIM18,932
OpenAI disclosures24
AI Hallucination Cases2,292
AIID1,732
MITRE ATLAS77
AVID1,745
AIAAIC2,196
NHTSA2
UK AISI1
FTC2

The collections include incident reports, hazards, court decisions, research exercises and regulatory proceedings.

THE CASE FOR SAFE FRONTIER AI

The future is worth
getting right.

We follow AI where it matters: the failures people need to understand and the progress worth imagining. Sources and context help visitors recognise risks, ask better questions and discover what is possible.

This website is powered by City Update, a Düsseldorf agency, working together with AI.

Explore the infrastructure behind AI
01 / RECOGNISE

Make the failure visible.

Record the actual impact, source evidence and uncertainty. Give affected people a voice.

02 / RESPOND

Show what changed.

Document containment, fixes and company responses. Separate action taken from claims made.

03 / LEARN

Earn confidence again.

Look for independent testing and follow-up evidence. Keep unresolved questions in view.

AI ALSO DEPENDS ON A PHYSICAL FOUNDATION
01

Chips

Specialist manufacturing and a global supply chain.

02

Electricity

Reliable power for training and everyday use.

03

Data centres

Physical facilities, networks and equipment.

04

Cooling

Heat management, with needs that vary by design.

05

People

Engineers, operators and accountable institutions.

A shared record. A stronger relationship.

Individuals, researchers and companies can submit evidence to this independently published register. The register has no regulatory authority and does not certify AI systems.

FOR THE PEOPLE INFORMING THE PUBLIC

A better-informed story.
A more informed public.

Good reporting makes room for both the promise of frontier AI and the consequences of its failures. Start with the evidence, ask what changed, and keep the unanswered questions visible.

01

Facts with their sources.

Provider statements, affected-party accounts and independent investigations, clearly distinguished.

02

The response is part of the story.

What was contained, what was fixed and what still needs evidence. Progress should be demonstrable.

03

Corrections deserve visibility.

A record that can be challenged and improved. Contact the editorial team for source questions or corrections.

Contact the press desk