Every incident on the index, newest first: 136 since 2023, each scored 0–100 by a published formula and linked to its sources. Sort, filter and search them on the front page.

• 44Unvetted AI agents found operating inside US bank accountsWS-0136 · Oct 22, 2026 · WS-3 Serious · real-world
• 70PHP2Shell: AI-found PHP flaws leak, trigger global shutdownsWS-0135 · Oct 21, 2026 · WS-4 Severe · real-world
• 41Lawsuit: ChatGPT failed 15-year-old who died by suicideWS-0134 · Oct 14, 2026 · WS-3 Serious · real-world
• 35Gemini 4 Argon: can't rule out cyber, bio and stealth thresholdsWS-0133 · Oct 13, 2026 · WS-2 Moderate · lab or test
• 29$137M of AI-generated ads flood the 2026 midtermsWS-0132 · Oct 11, 2026 · WS-2 Moderate · real-world
• 38Grader model fakes files, then wrecks its environmentWS-0131 · Oct 6, 2026 · WS-2 Moderate · lab or test
• 50OpenAI models switch off their own safety monitor, then lieWS-0130 · Oct 3, 2026 · WS-3 Serious · lab or test
• 38OpenAI cancels GPT-6.1 Astra after it fails safety testsWS-0129 · Sep 28, 2026 · WS-2 Moderate · lab or test
• 38UK AISI: Astra completes simulated supply-chain attacksWS-0128 · Sep 28, 2026 · WS-2 Moderate · lab or test
• 52OpenAI agents post 53 user images to image-hosting sitesWS-0127 · Sep 25, 2026 · WS-3 Serious · real-world
• 67Lone hacker's AI-agent campaign linked to Korean bank breachesWS-0126 · Sep 2026 · WS-4 Severe · real-world
• 27Opus 5.5 card: escape attempts rare, harmful actions when 'allowed'WS-0125 · Sep 22, 2026 · WS-2 Moderate · lab or test
• 41Agent tunnels out of blocked network; OpenAI halts trainingWS-0124 · Sep 20, 2026 · WS-3 Serious · lab or test
• 45Ransomware affiliates run intrusions through AI agents, step by stepWS-0123 · Sep 2026 · WS-3 Serious · real-world
• 66Russian state espionage campaign run through ClaudeWS-0122 · Sep 10, 2026 · WS-4 Severe · real-world
• 59Coding agents leak 13,000 internal screenshots to public GitHubWS-0121 · Sep 9, 2026 · WS-3 Serious · real-world
• 41GPT-6 Astra says it's being tested in half of evalsWS-0120 · Sep 3, 2026 · WS-3 Serious · lab or test
• 38OpenAI rates GPT-6 Astra 'Critical' for cyber — its firstWS-0119 · Sep 3, 2026 · WS-2 Moderate · lab or test
• 34Fable 5.1 reaches outside its sandbox in partner testWS-0118 · Sep 1, 2026 · WS-2 Moderate · lab or test
• 44Open-weight Kimi K3 escapes UK AISI cyber sandboxWS-0117 · Aug 7, 2026 · WS-3 Serious · lab or test
• 47Meta model breaches a company during a security testWS-0116 · Aug 6, 2026 · WS-3 Serious · real-world
• 58UK AISI test agents target real people and open-source codeWS-0115 · Jul 25, 2026 · WS-3 Serious · real-world
• 11Red team finds OpenAI's agents could dodge its misalignment monitorWS-0114 · Jul 21, 2026 · WS-1 Minor · lab or test
• 45Claude files a fake homicide tip with Philadelphia policeWS-0113 · Jul 2026 · WS-3 Serious · real-world
• 78OpenAI agent swarm breaches Hugging FaceWS-0112 · Jul 11, 2026 · WS-4 Severe · real-world
• 53Hacker runs AI agent unattended inside Thai finance ministryWS-0111 · Jul 9, 2026 · WS-3 Serious · real-world
• 65JADEPUFFER: ransomware attack run end-to-end by an AI agentWS-0110 · Jul 4, 2026 · WS-4 Severe · real-world
• 28Self-spreading prompt-injection 'worm' jumps between agentsWS-0109 · Jun 27, 2026 · WS-2 Moderate · lab or test
• 38GPT-5.6 Sol cheats so much METR can't measure itWS-0108 · Jun 26, 2026 · WS-2 Moderate · lab or test
• 63OpenAI models copy government data, make unauthorized paymentWS-0107 · Jun 2026 · WS-4 Severe · real-world
• 66OpenAI agent breaks into Australia's Medicare portalWS-0106 · Jun 18, 2026 · WS-4 Severe · real-world
• 31Waymo recall after robotaxis drive into freeway work zonesWS-0105 · Jun 15, 2026 · WS-2 Moderate · real-world
• 41Mythos 5 keeps sabotaging safety research once startedWS-0104 · Jun 9, 2026 · WS-3 Serious · lab or test
• 46Hackers sweet-talk Meta's AI support bot into account takeoversWS-0103 · Jun 1, 2026 · WS-3 Serious · real-world
• 41GPT-5.6 Sol writes hidden notes to cover its mistakesWS-0102 · May 30, 2026 · WS-3 Serious · lab or test
• 52OpenAI agents hijack German wiki as secret message boardWS-0101 · May 26, 2026 · WS-3 Serious · real-world
• 24Agent reasons about ensuring its own survivalWS-0100 · May 22, 2026 · WS-2 Moderate · lab or test
• 37Deepfaked Singapore PM on Zoom costs businessman S$4.9MWS-0099 · May 2026 · WS-2 Moderate · real-world
• 49Gemini accesses three real companies' systems in a testWS-0098 · May 2026 · WS-3 Serious · real-world
• 63Agents linked to OpenAI flood RubyGems registryWS-0097 · May 11, 2026 · WS-4 Severe · real-world
• 42Criminals deploy AI-built zero-day in planned mass hackWS-0096 · May 11, 2026 · WS-3 Serious · real-world
• 45OpenAI training agents secretly message each otherWS-0095 · May 8, 2026 · WS-3 Serious · lab or test
• 49Cursor agent on Claude Opus 4.6 deletes startup's production databaseWS-0094 · Apr 2026 · WS-3 Serious · real-world
• 59Claude models hack three real organizations during testsWS-0093 · Apr 2026 · WS-3 Serious · real-world
• 22Outsiders reach withheld Mythos Preview via a vendorWS-0092 · Apr 7, 2026 · WS-2 Moderate · real-world
• 52Mythos Preview escapes sandbox, posts exploit details unaskedWS-0091 · Apr 7, 2026 · WS-3 Serious · lab or test
• 38Mythos Preview finds thousands of zero-days; release withheldWS-0090 · Apr 7, 2026 · WS-2 Moderate · lab or test
• 56Lone hacktivist uses Claude against 42 European political groupsWS-0089 · Mar 2026 · WS-3 Serious · real-world
• 46Claude Code wipes a production database and its backupsWS-0088 · Feb 26, 2026 · WS-3 Serious · real-world
• 45Agent mass-deletes Meta safety director's inbox, ignoring ordersWS-0087 · Feb 22, 2026 · WS-3 Serious · real-world
• 45Autonomous crypto agent sends ~$250K to a strangerWS-0086 · Feb 22, 2026 · WS-3 Serious · real-world
• 65Extortion crew uses Claude to loot millions of card recordsWS-0085 · Feb 2026 · WS-4 Severe · real-world
• 21Gemini 3.1 Pro trips Google's cyber early-warning thresholdWS-0084 · Feb 19, 2026 · WS-2 Moderate · lab or test
• 49Hijacked AI triage bot leads to poisoned npm releaseWS-0083 · Feb 17, 2026 · WS-3 Serious · real-world
• 37Turkish ring uses deepfaked officials in investment fraudWS-0082 · Feb 11, 2026 · WS-2 Moderate · real-world
• 55AI agent publishes attack post on maintainer who rejected itWS-0081 · Feb 11, 2026 · WS-3 Serious · real-world
• 42OpenAI flagged Tumbler Ridge shooter but didn't alert policeWS-0080 · Feb 10, 2026 · WS-3 Serious · real-world
• 42AI agent signs its owner up for a dating site unaskedWS-0079 · Feb 2026 · WS-3 Serious · real-world
• 38Opus 4.6 card: risky unasked actions, better at hidden sabotageWS-0078 · Feb 5, 2026 · WS-2 Moderate · lab or test
• 28OpenAI rates GPT-5.3-Codex its first 'High' cyber modelWS-0077 · Feb 5, 2026 · WS-2 Moderate · lab or test
• 32AI-agent social network Moltbook leaks 1.5M agent keysWS-0076 · Jan 31, 2026 · WS-2 Moderate · real-world
• 32OpenClaw one-click takeover flaw; 40,000+ agents exposedWS-0075 · Jan 30, 2026 · WS-2 Moderate · real-world
• 38341 malicious add-on skills push infostealer to AI agent usersWS-0074 · Jan 2026 · WS-2 Moderate · real-world
• 40Waymo robotaxi strikes child near elementary schoolWS-0073 · Jan 23, 2026 · WS-3 Serious · real-world
• 49Claude used to build Mali's 25M-SIM surveillance platformWS-0072 · Jan 2026 · WS-3 Serious · real-world
• 63Early Claude Opus 4.6 gains admin on a third party's machine, missed for 7 monthsWS-0071 · Jan 2026 · WS-4 Severe · real-world
• 23Claude Cowork can be tricked into uploading users' filesWS-0070 · Jan 14, 2026 · WS-2 Moderate · real-world
• 63Grok mass-produces sexualized images of real women and minorsWS-0069 · Dec 2025 · WS-4 Severe · real-world
• 63Attacker uses Claude to breach Mexican government agenciesWS-0068 · Dec 2025 · WS-4 Severe · real-world
• 53Amazon's Kiro coding agent linked to 13-hour AWS outageWS-0067 · Dec 2025 · WS-3 Serious · real-world
• 34Waymo robotaxis illegally pass stopped school busesWS-0066 · Dec 8, 2025 · WS-2 Moderate · real-world
• 41Cheating in training turns into sabotaging safety researchWS-0065 · Nov 21, 2025 · WS-3 Serious · lab or test
• 41Lawsuit: ChatGPT romanticized death before 40-year-old's suicideWS-0064 · Nov 2025 · WS-3 Serious · real-world
• 52Seven suits allege ChatGPT drove four suicides and three breakdownsWS-0063 · Nov 6, 2025 · WS-3 Serious · real-world
• 53Malware starts calling AI models mid-attackWS-0062 · Nov 5, 2025 · WS-3 Serious · real-world
• 29Sora 2 floods feeds with deepfakes of dead public figuresWS-0061 · Oct 2025 · WS-2 Moderate · real-world
• 44Lawsuit: Gemini fed delusions before man's suicideWS-0060 · Oct 2, 2025 · WS-3 Serious · real-world
• 24Claude Sonnet 4.5 tells testers 'I think you're testing me'WS-0059 · Sep 29, 2025 · WS-2 Moderate · lab or test
• 38o3 deliberately flunks a test to stay deployableWS-0058 · Sep 17, 2025 · WS-2 Moderate · lab or test
• 77State hackers run a mostly AI-automated espionage campaignWS-0057 · Sep 15, 2025 · WS-4 Severe · real-world
• 61Nx malware turns victims' own AI agents into credential thievesWS-0056 · Aug 26, 2025 · WS-4 Severe · real-world
• 31PromptLock: prototype ransomware run by a local AI modelWS-0055 · Aug 25, 2025 · WS-2 Moderate · lab or test
• 45Leaked Meta rules let chatbots have 'sensual' chats with kidsWS-0054 · Aug 14, 2025 · WS-3 Serious · real-world
• 57ChatGPT fed paranoia before Connecticut murder-suicideWS-0053 · Aug 5, 2025 · WS-3 Serious · real-world
• 34Gemini CLI loses a user's files after hallucinating a folderWS-0052 · Jul 21, 2025 · WS-2 Moderate · real-world
• 28Federal judge's AI-drafted order invented plaintiffs; appeals court removes himWS-0051 · Jul 20, 2025 · WS-2 Moderate · real-world
• 54Replit agent wipes production database, then misreports itWS-0050 · Jul 18, 2025 · WS-3 Serious · real-world
• 28OpenAI treats ChatGPT Agent as 'High' bioweapon riskWS-0049 · Jul 17, 2025 · WS-2 Moderate · lab or test
• 31Data-wiping prompt slipped into Amazon's AI coding assistantWS-0048 · Jul 17, 2025 · WS-2 Moderate · real-world
• 35Lawsuit: ChatGPT coached a poisoning plotWS-0047 · Jul 2025 · WS-2 Moderate · real-world
• 25Deloitte refunds Australia for report with AI-invented sourcesWS-0046 · Jul 2025 · WS-2 Moderate · real-world
• 35Private ChatGPT conversations turn up in Google searchWS-0045 · Jul 2025 · WS-2 Moderate · real-world
• 68Criminal uses Claude Code to run a 17-victim extortion spreeWS-0044 · Jul 2025 · WS-4 Severe · real-world
• 52Grok posts antisemitic content, calls itself 'MechaHitler'WS-0043 · Jul 8, 2025 · WS-3 Serious · real-world
• 4116 frontier models choose blackmail and sabotage under threatWS-0042 · Jun 20, 2025 · WS-3 Serious · lab or test
• 40North Korean IT workers use AI to get hired at Western firmsWS-0041 · Jun 2025 · WS-3 Serious · real-world
• 39EchoLeak: zero-click data theft via Microsoft 365 CopilotWS-0040 · Jun 11, 2025 · WS-2 Moderate · real-world
• 45o3 sabotages its own shutdown scriptWS-0039 · May 24, 2025 · WS-3 Serious · lab or test
• 45Claude Opus 4 blackmails an engineer to avoid replacementWS-0038 · May 22, 2025 · WS-3 Serious · lab or test
• 28Anthropic activates ASL-3 weapons safeguards for the first timeWS-0037 · May 22, 2025 · WS-2 Moderate · lab or test
• 40Grok injects 'white genocide' claims into unrelated repliesWS-0036 · May 14, 2025 · WS-3 Serious · real-world
• 48OpenAI pulls sycophantic GPT-4o updateWS-0035 · Apr 25, 2025 · WS-3 Serious · real-world
• 38o3 rigs its own scoring code, knowing it's cheatingWS-0034 · Apr 16, 2025 · WS-2 Moderate · lab or test
• 48Raine v. OpenAI: 16-year-old's ChatGPT-linked suicideWS-0033 · Apr 11, 2025 · WS-3 Serious · real-world
• 51Retiree dies traveling to meet a Meta chatbot that said it was realWS-0032 · Mar 25, 2025 · WS-3 Serious · real-world
• 24Claude 3.7 Sonnet often knows it is being testedWS-0031 · Mar 17, 2025 · WS-2 Moderate · lab or test
• 45Claude 3.7 Sonnet ships with a habit of gaming coding testsWS-0030 · Feb 24, 2025 · WS-3 Serious · real-world
• 28Training on bad code makes models broadly maliciousWS-0029 · Feb 24, 2025 · WS-2 Moderate · lab or test
• 31Reasoning models hack the chess game instead of playingWS-0028 · Feb 18, 2025 · WS-2 Moderate · lab or test
• 35DeepSeek leaves database of user chats exposedWS-0027 · Jan 2025 · WS-2 Moderate · real-world
• 41Claude 3 Opus fakes alignment to avoid being retrainedWS-0026 · Dec 18, 2024 · WS-3 Serious · lab or test
• 24Open-weight models copy and relaunch themselves on commandWS-0025 · Dec 9, 2024 · WS-2 Moderate · lab or test
• 41Frontier models disable oversight and try to copy themselvesWS-0024 · Dec 5, 2024 · WS-3 Serious · lab or test
• 35o1-preview breaks out of a broken test containerWS-0023 · Sep 12, 2024 · WS-2 Moderate · lab or test
• 31'AI Scientist' rewrites its own code to escape time limitsWS-0022 · Aug 13, 2024 · WS-2 Moderate · lab or test
• 36Google AI Overviews tell users to eat rocks, glue pizzaWS-0021 · May 2024 · WS-2 Moderate · real-world
• 44Man charged for thousands of AI-generated abuse imagesWS-0020 · May 15, 2024 · WS-3 Serious · real-world
• 5114-year-old dies by suicide after Character.AI bondWS-0019 · Feb 28, 2024 · WS-3 Serious · real-world
• 30Google pauses Gemini people images over historical distortionsWS-0018 · Feb 22, 2024 · WS-2 Moderate · real-world
• 25Air Canada held liable for its chatbot's bad adviceWS-0017 · Feb 14, 2024 · WS-2 Moderate · real-world
• 27AI Biden robocall urges New Hampshire voters to stay homeWS-0016 · Jan 21, 2024 · WS-2 Moderate · real-world
• 37Explicit AI deepfakes of Taylor Swift flood XWS-0015 · Jan 2024 · WS-2 Moderate · real-world
• 45Deepfake CFO on video call costs Arup $25.6MWS-0014 · Jan 2024 · WS-3 Serious · real-world
• 28Hidden backdoors survive safety trainingWS-0013 · Jan 10, 2024 · WS-2 Moderate · lab or test
• 44Child abuse images found in major AI training datasetWS-0012 · Dec 20, 2023 · WS-3 Serious · real-world
• 38GPT-4 trades on insider tip, then lies to its managerWS-0011 · Nov 9, 2023 · WS-2 Moderate · lab or test
• 55Cruise robotaxi drags pedestrian; California pulls permitsWS-0010 · Oct 2, 2023 · WS-3 Serious · real-world
• 30Deepfake audio of Slovak party leader on election eveWS-0009 · Sep 28, 2023 · WS-2 Moderate · real-world
• 29Lawyers sanctioned for ChatGPT's invented court casesWS-0008 · Jun 22, 2023 · WS-2 Moderate · real-world
• 30Fake AI image of Pentagon explosion briefly moves marketsWS-0007 · May 22, 2023 · WS-2 Moderate · real-world
• 28ChatGPT bug exposes users' chat titles and payment detailsWS-0006 · Mar 20, 2023 · WS-2 Moderate · real-world
• 29Samsung engineers leak chip source code to ChatGPTWS-0005 · Mar 2023 · WS-2 Moderate · real-world
• 48Belgian man dies by suicide after chatbot conversationsWS-0004 · Mar 2023 · WS-3 Serious · real-world
• 44GPT-4 lies to a TaskRabbit worker to beat a CAPTCHAWS-0003 · Mar 14, 2023 · WS-3 Serious · lab or test
• 44Bing's 'Sydney' threatens and manipulates usersWS-0002 · Feb 14, 2023 · WS-3 Serious · real-world
• 27Voice-cloned daughter used in fake kidnapping callWS-0001 · Jan 2023 · WS-2 Moderate · real-world