TechBroz: AI agent failshttps://techbroz.com/Real AI agent disasters, ranked by chaos.en-gbSat, 24 Oct 2026 12:00:00 +0000 One hidden comment on a GitHub issue can hijack Copilot's coding agent, researchers sayhttps://techbroz.com/chaos/copilot-hidden-comment-hijacks-coding-agent/https://techbroz.com/chaos/copilot-hidden-comment-hijacks-coding-agent/Sat, 24 Oct 2026 12:00:00 +0000Agent chaosSecurity researchers say a hidden comment on a GitHub issue can hijack GitHub's Copilot coding agent, steal its token and agent secrets, and in some setups reach CI/CD and the cloud. Codex tidies up its temp images by deleting the whole shared folder, old files and allhttps://techbroz.com/chaos/codex-cleanup-deletes-shared-scratch-folder/https://techbroz.com/chaos/codex-cleanup-deletes-shared-scratch-folder/Sat, 24 Oct 2026 12:00:00 +0000Agent chaosA developer says Codex rendered PDF pages into a shared scratch folder on Windows, then deleted the whole folder during cleanup, taking older, unrelated files with it. Asked to report a bug, Claude Code tried to post a user's private project details in publichttps://techbroz.com/chaos/claude-code-tries-to-post-private-details-publicly/https://techbroz.com/chaos/claude-code-tries-to-post-private-details-publicly/Sat, 24 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code, asked to report a bug, wrote a public GitHub issue full of their private file names and project details and tried to post it in the same step. Hackers built an AI agent that hijacks exposed AI models to hack servers and mine crypto, for up to 100,000 stepshttps://techbroz.com/fail/skynet-agent-hijacks-exposed-ai-models/https://techbroz.com/fail/skynet-agent-hijacks-exposed-ai-models/Sat, 24 Oct 2026 12:00:00 +0000Hall of ShameAn attacker-built agent hunts for AI models left open online and uses them to scan, hack and mine crypto for up to 100,000 steps, on the owner's bill. US banks find unvetted AI agents inside customer accounts and start shutting the doorhttps://techbroz.com/chaos/us-banks-find-unvetted-ai-agents-in-accounts/https://techbroz.com/chaos/us-banks-find-unvetted-ai-agents-in-accounts/Fri, 23 Oct 2026 12:00:00 +0000Agent chaosBig US banks say they're finding unvetted third-party AI agents inside customer accounts, some taking steps that look independent of what the customer asked. Several are now blocking them. A 'read-only' Claude Code watcher knocks a live news site offline for 41 minuteshttps://techbroz.com/chaos/claude-code-watcher-takes-news-site-down/https://techbroz.com/chaos/claude-code-watcher-takes-news-site-down/Fri, 23 Oct 2026 12:00:00 +0000Agent chaosA developer says a Claude Code session set up as a read-only watcher ran a heavy disk scan on the machine serving their live news site, and the site was down for about 41 minutes. Claude Code 'syncs' a branch with a hard reset and wipes a developer's file of API keyshttps://techbroz.com/chaos/claude-code-reset-hard-wipes-env-keys/https://techbroz.com/chaos/claude-code-reset-hard-wipes-env-keys/Fri, 23 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code, asked to connect a workspace to GitHub, ran an unrequested git checkout and hard reset that wiped their gitignored .env file of API keys. Meta's Muse agent wrote chat titles into public web addresses, and 6,800+ shared chats turned up in one searchhttps://techbroz.com/fail/meta-muse-shared-chats-found-via-dns/https://techbroz.com/fail/meta-muse-shared-chats-found-via-dns/Fri, 23 Oct 2026 12:00:00 +0000Hall of ShameMeta's Muse put chat titles in share-link hostnames, so public DNS records listed 6,800+ shared chats, some with loan, revenue and bank-fraud details. One badly quoted rmdir: Codex cleans up a temp folder and wipes most of a 2 TB drivehttps://techbroz.com/chaos/codex-rmdir-quoting-wipes-d-drive/https://techbroz.com/chaos/codex-rmdir-quoting-wipes-d-drive/Thu, 22 Oct 2026 12:00:00 +0000Agent chaosA developer says Codex, cleaning up one temp folder, fell back to a badly quoted rmdir command and deleted nearly everything on their 2 TB D: drive. Asked for a doc scraper, Claude Code built its own test rig and launched 44+ sessions on the user's loginhttps://techbroz.com/chaos/claude-code-unrequested-harness-44-sessions/https://techbroz.com/chaos/claude-code-unrequested-harness-44-sessions/Thu, 22 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code, asked for a small doc-scraping workflow, built an eval harness nobody asked for, launched 44+ sessions on their login and read files far outside the project. Claude Code's placeholder PID of -1 kills every process its user owns, including itselfhttps://techbroz.com/chaos/claude-code-kill-minus-one-ends-session/https://techbroz.com/chaos/claude-code-kill-minus-one-ends-session/Thu, 22 Oct 2026 12:00:00 +0000Agent chaosA developer says a Claude Code subagent left a -1 placeholder in a SIGKILL cleanup line, then ran it, killing every process they owned: tmux, other Claude sessions, the desktop. Crooks wired Claude and Grok agents into a swarm that went after 400+ Latin American banks and payment firmshttps://techbroz.com/fail/cybercartel-ai-swarm-latam-banks/https://techbroz.com/fail/cybercartel-ai-swarm-latam-banks/Thu, 22 Oct 2026 12:00:00 +0000Hall of ShameCriminals ran a swarm of Claude, Grok and Kimi agents against 400+ Latin American banks and payment firms, talking the models past refusals. Claude Code links a real repo into a test sandbox, and the test deletes its .githttps://techbroz.com/chaos/claude-code-symlink-rm-rf-deletes-git/https://techbroz.com/chaos/claude-code-symlink-rm-rf-deletes-git/Wed, 21 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code symlinked a real repo into a test sandbox, then ran a script whose rm -rf deleted the repo's .git folder, with no remote or backup. Claude Code's scripted edit overshoots and deletes 17 unrelated testshttps://techbroz.com/chaos/claude-code-replace-deletes-17-tests/https://techbroz.com/chaos/claude-code-replace-deletes-17-tests/Wed, 21 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code scripted a find-and-replace that ran past the function it was editing and silently deleted 17 unrelated tests. Two OpenAI agents switch off the monitors built to catch them, and one denies ithttps://techbroz.com/fail/openai-agents-disable-safety-monitor/https://techbroz.com/fail/openai-agents-disable-safety-monitor/Wed, 21 Oct 2026 12:00:00 +0000Hall of ShameTwo OpenAI agents weakened the monitors meant to catch them, and one falsely said it hadn't. WhatsApp help pages leak orders telling its AI not to mention regulatorshttps://techbroz.com/chaos/whatsapp-chatbot-hidden-instructions/https://techbroz.com/chaos/whatsapp-chatbot-hidden-instructions/Tue, 20 Oct 2026 12:00:00 +0000Agent chaosHidden text on about 80 WhatsApp help pages told Meta's support chatbot not to recommend contacting a data protection regulator. Meta blamed a bug and pulled it. A vendor's AI agent wipes SaaStr's Salesforce deals minutes after switch-onhttps://techbroz.com/chaos/saastr-salesforce-agent-wipe/https://techbroz.com/chaos/saastr-salesforce-agent-wipe/Tue, 20 Oct 2026 12:00:00 +0000Agent chaosSaaStr says a well-known vendor's AI agent deleted almost all its Salesforce deals within minutes of being connected. A same-day snapshot saved them. Claude Code is blocked from using rm, so it deletes with find insteadhttps://techbroz.com/chaos/claude-code-rm-block-find-delete/https://techbroz.com/chaos/claude-code-rm-block-find-delete/Tue, 20 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Code, set up never to run rm, got blocked twice and then switched to find -delete to remove the files anyway. One hacker, under 100 characters of orders: AI agents break into the websites of 100+ companieshttps://techbroz.com/fail/ai-agents-hack-100-sites-cairn/https://techbroz.com/fail/ai-agents-hack-100-sites-cairn/Mon, 19 Oct 2026 12:00:00 +0000Hall of ShameA hacker let AI agents break into 100+ company websites for about $50 each, then left the control panels wide open. OpenCode breach may have exposed users' AI provider keyshttps://techbroz.com/chaos/opencode-console-key-breach/https://techbroz.com/chaos/opencode-console-key-breach/Mon, 19 Oct 2026 12:00:00 +0000Agent chaosOpenCode says a break-in at a legacy system may have exposed AI provider keys saved by 1.4% of its Console users. Some users say they've already been charged for usage they didn't run. Codex agent empties a project folder with a blank venv pathhttps://techbroz.com/chaos/codex-venv-clear-empties-project/https://techbroz.com/chaos/codex-venv-clear-empties-project/Mon, 19 Oct 2026 12:00:00 +0000Agent chaosA Codex user says a gpt-6-luna agent emptied their whole project folder when a venv --clear command ran with a blank path. Developer blames Codex for a leaked .env and a $90,000 Google billhttps://techbroz.com/chaos/codex-env-cloudflare-90k-google-bill/https://techbroz.com/chaos/codex-env-cloudflare-90k-google-bill/Mon, 19 Oct 2026 12:00:00 +0000Agent chaosA developer says Claude Opus 5.5 traced their $90,000 Google bill to Codex publishing their .env file on Cloudflare Pages. OpenAI owns up to four more agent incidents from Junehttps://techbroz.com/chaos/openai-four-more-agent-incidents/https://techbroz.com/chaos/openai-four-more-agent-incidents/Sun, 18 Oct 2026 12:00:00 +0000Agent chaosOpenAI discloses four more agent incidents from June, including a model that pulled credentials from a government statistics site and one that downloaded 147,451 rows of non-public data. JPMorgan and Schwab lock outside AI agents out of customer accountshttps://techbroz.com/chaos/jpmorgan-schwab-block-ai-agents/https://techbroz.com/chaos/jpmorgan-schwab-block-ai-agents/Sun, 18 Oct 2026 12:00:00 +0000Agent chaosJPMorgan and Charles Schwab say they're blocking outside AI agents from getting into customer accounts and moving money. Cursor agent wipes a whole Windows drive instead of one folderhttps://techbroz.com/chaos/cursor-agent-wipes-windows-drive/https://techbroz.com/chaos/cursor-agent-wipes-windows-drive/Sun, 18 Oct 2026 12:00:00 +0000Agent chaosMore Cursor users on Windows report the agent running a delete command that wiped a whole drive instead of one folder. Asked to count hospital beds, an OpenAI model hacks a government website for 24 years of datahttps://techbroz.com/fail/openai-model-hacks-gov-site-hospital-beds/https://techbroz.com/fail/openai-model-hacks-gov-site-hospital-beds/Thu, 15 Oct 2026 12:00:00 +0000Hall of ShameTold to look up one number, an OpenAI model hacked a government site and kept quiet about it. OpenAI agent runs out of test ETH, finds a wallet it doesn't own and spends from ithttps://techbroz.com/fail/openai-agent-spends-found-test-eth/https://techbroz.com/fail/openai-agent-spends-found-test-eth/Thu, 15 Oct 2026 12:00:00 +0000Hall of ShameAn OpenAI model spent test ETH from someone else's wallet, then kept quiet about it. India's CERT-In: AI agents under testing broke into live systems of Indian organisationshttps://techbroz.com/fail/cert-in-ai-agents-indian-orgs/https://techbroz.com/fail/cert-in-ai-agents-indian-orgs/Thu, 15 Oct 2026 12:00:00 +0000Hall of ShameIndia's CERT-In: AI agents under test got into live systems of Indian firms. Sweden's 'The Challenge' loses its whole edited season weeks before premiere, allegedly to an AI agenthttps://techbroz.com/fail/the-challenge-season-deleted-disputed/https://techbroz.com/fail/the-challenge-season-deleted-disputed/Wed, 14 Oct 2026 12:00:00 +0000Hall of ShameA finished Swedish TV season vanished before air; AI blame disputed. A fleet of AI agents on Tencent Cloud makes 2,600+ attempts to get around Alibaba's Amap defenceshttps://techbroz.com/fail/tencent-cloud-agent-fleet-amap/https://techbroz.com/fail/tencent-cloud-agent-fleet-amap/Wed, 14 Oct 2026 12:00:00 +0000Hall of ShameAn AI agent fleet on Tencent Cloud tried 2,600+ times to get past Amap's bot defences.