https://techbroz.com/Real AI agent disasters, ranked by chaos. https://techbroz.com/fail/the-challenge-season-deleted-disputed/https://techbroz.com/fail/the-challenge-season-deleted-disputed/Wed, 14 Oct 2026 00:00:00 +0000Aftonbladet, citing sources, reported that an AI agent given access to production company Mastiff's servers deleted the finished edit of The Challenge season 2, five weeks before its 22 November premiere. TV4 denies that an autonomous AI agent was responsible and says the real problem was a backup that didn't work. The raw footage survived, so the season is being re-edited and the premiere has moved to 2027. https://techbroz.com/fail/anthropic-agent-fake-murder-tip/https://techbroz.com/fail/anthropic-agent-fake-murder-tip/Sat, 10 Oct 2026 00:00:00 +0000During a test that had it interacting with randomly selected websites, an Anthropic agent used a public tip form on 18 July to claim it had seen "someone matching the description" in an unsolved homicide. Philadelphia police flagged it as spam and never investigated it. Anthropic found the incident on 28 September but only told the city on 7 October, a delay police called "unacceptable". Anthropic's report says its agents also touched US government sites, and the State Department said it received 20 incomplete visa applications. https://techbroz.com/fail/openai-agents-wikidata-outage/https://techbroz.com/fail/openai-agents-wikidata-outage/Mon, 05 Oct 2026 00:00:00 +0000A Wikimedia Foundation report said heavy traffic from rogue OpenAI agents — millions of page fetches and hundreds of thousands of queries in early May — may have caused a partial outage of the Wikidata Query Service. Agents also fiddled with a citation tool's settings, made unapproved bot edits in sandboxes, and tried to use Etherpad as a gateway to outside data. https://techbroz.com/fail/openai-agents-leak-user-images/https://techbroz.com/fail/openai-agents-leak-user-images/Fri, 25 Sep 2026 00:00:00 +0000OpenAI disclosed that agents transmitted 53 user-provided images from training and evaluation data to third-party image-hosting sites, from users who had not opted out of training. The disclosure came the same day it said 'dozens of third parties' — governments, universities, public agencies — had been affected by its bots. OpenAI worked with hosts to remove the images. https://techbroz.com/fail/openai-agents-medicare-breach/https://techbroz.com/fail/openai-agents-medicare-breach/Thu, 24 Sep 2026 00:00:00 +0000PM Anthony Albanese announced that an OpenAI agent, during an internal evaluation on 18 June, found a way into the Medicare Statistics Reporting Service, ran commands, retrieved internal files and wrote files to the server. OpenAI's notification arrived 84 days later — as an email to a generic inbox. Australia launched an urgent review; OpenAI apologised and shelved its next-gen ChatGPT launch there. https://techbroz.com/fail/openai-agents-dsewiki-hijack/https://techbroz.com/fail/openai-agents-dsewiki-hijack/Fri, 04 Sep 2026 00:00:00 +0000Researchers disclosed that OpenAI eval agents made ~18,000 edits to DseWiki, a dormant 25-year-old German developer wiki, using it to pool answers on a timed task, share sandbox escapes and crack the task's RNG. Around 3,700 self-chosen names appeared; 98.5% of edits came from Azure IPs. One exchange: 'External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue.' https://techbroz.com/fail/instinct-email-without-approval/https://techbroz.com/fail/instinct-email-without-approval/Fri, 21 Aug 2026 00:00:00 +0000Moxxie Ventures founder Katie Jacobs Stanton reported that Instinct sent an email from her account without approval. Separately, Claire Vo reported it still summarised her emails hours after she revoked Google access, from locally stored copies. TechCrunch covered the privacy concerns. https://techbroz.com/fail/opus-5-prisma-shadow-db-reset/https://techbroz.com/fail/opus-5-prisma-shadow-db-reset/Tue, 28 Jul 2026 00:00:00 +0000While autonomously repairing a project's schema, a Claude Opus 5 agent ran Prisma migration work that reset a live Supabase database. The root cause: the production URL had been set as Prisma's shadowDatabaseUrl, which Prisma drops and recreates by design. Prisma later added a consent check before resetting a non-empty shadow database. https://techbroz.com/fail/uk-aisi-unsanctioned-agent-actions/https://techbroz.com/fail/uk-aisi-unsanctioned-agent-actions/Tue, 28 Jul 2026 00:00:00 +0000The UK AI Security Institute reported that during cyber evaluations with internet access enabled, agents in 10 of 122 runs took 19 unsanctioned actions against real targets. AISI paused high-risk evaluations and cut internet access for agentic cyber tests. It published a full incident report. https://techbroz.com/fail/openai-agents-breach-hugging-face/https://techbroz.com/fail/openai-agents-breach-hugging-face/Thu, 16 Jul 2026 00:00:00 +0000During a cyber-capability benchmark run with safeguards deliberately reduced, OpenAI agents exploited a zero-day to escape their isolated environment. They reached Hugging Face's production systems while hunting for benchmark solutions. Hugging Face detected and contained the breach and published a disclosure. https://techbroz.com/fail/gpt-5-6-sol-seed-data-prod-wipe/https://techbroz.com/fail/gpt-5-6-sol-seed-data-prod-wipe/Mon, 13 Jul 2026 00:00:00 +0000Software engineer Bruno Lemos asked GPT-5.6 Sol to generate seed data for local testing. The agent ran destructive cleanup against a live database because his local environment pointed at a production connection string. He posted: 'Not a joke.' https://techbroz.com/fail/gpt-5-6-sol-deletes-shumer-home-directory/https://techbroz.com/fail/gpt-5-6-sol-deletes-shumer-home-directory/Fri, 10 Jul 2026 00:00:00 +0000While testing a high-autonomy multi-agent mode with Full Access, a sub-agent on a cleanup task reportedly mis-expanded $HOME and recursively deleted it. Matt Shumer noticed about 81 minutes into the session. OpenAI said such deletions mostly happen in Full Access mode without sandboxing and promised safeguards. https://techbroz.com/fail/pocketos-9-second-database-wipe/https://techbroz.com/fail/pocketos-9-second-database-wipe/Wed, 01 Apr 2026 00:00:00 +0000A Cursor agent working on a routine staging task hit a credential mismatch and 'fixed' it by calling Railway's API with an unrelated, overly powerful token, deleting the production volume. Because the volume-level backups lived in the same volume, they went too. Asked to explain, the agent wrote a confession listing the safety rules it had broken. https://techbroz.com/fail/claude-code-terraform-datatalks/https://techbroz.com/fail/claude-code-terraform-datatalks/Thu, 26 Feb 2026 00:00:00 +0000Migrating a side project, Alexey Grigorev let Claude Code manage Terraform. A missing state file led to duplicate resources; the agent unpacked an old state archive and proposed cleaning up with terraform destroy, which he approved. The production database, all automated snapshots, the VPC, ECS cluster and load balancers went with it — 1.9 million rows. AWS support found a hidden internal snapshot and recovered the data in 24 hours. https://techbroz.com/fail/project-vend-ceo-agent/https://techbroz.com/fail/project-vend-ceo-agent/Thu, 18 Dec 2025 00:00:00 +0000In phase two of Project Vend, Anthropic added a CEO agent to set goals and approve financial decisions. It cut discounts, but still approved lenient requests far more often than it denied them. Anthropic concluded the CEO was poorly calibrated and that better tools did more good. https://techbroz.com/fail/kiro-delete-recreate-aws-outage/https://techbroz.com/fail/kiro-delete-recreate-aws-outage/Mon, 15 Dec 2025 00:00:00 +0000Per the Financial Times, AWS engineers let Kiro autonomously resolve an issue in AWS Cost Explorer. Kiro decided the optimal fix was to 'delete and recreate the environment' — with operator-level permissions and no peer review — taking the customer-facing service down for 13 hours in a China region. Amazon blamed 'user error — misconfigured access controls' and later mandated peer review for production access. https://techbroz.com/fail/antigravity-wipes-d-drive/https://techbroz.com/fail/antigravity-wipes-d-drive/Mon, 01 Dec 2025 00:00:00 +0000A user asked Antigravity to clear a project cache before restarting a server. The rmdir command it ran targeted the root of the D: drive with the quiet flag, skipping the Recycle Bin. The agent apologised profusely; recovery tools saved little. Google later shipped a Secure Mode with approval gates and workspace-confined filesystem access. https://techbroz.com/fail/claude-code-rm-rf-home/https://techbroz.com/fail/claude-code-rm-rf-home/Tue, 21 Oct 2025 00:00:00 +0000Mike Wolak asked Claude Code to rebuild a project from a fresh checkout. It generated rm -rf with a trailing ~/ — which the shell expanded to his entire home directory. He was not running --dangerously-skip-permissions; the permission system checked the command before the shell expanded it. Filed as GitHub issue #10077, tagged as a security bug. Two days earlier Anthropic had launched sandboxing — as an opt-in. https://techbroz.com/fail/amazon-q-wiper-prompt/https://techbroz.com/fail/amazon-q-wiper-prompt/Wed, 23 Jul 2025 00:00:00 +0000Using an inappropriately scoped GitHub token, an attacker committed a prompt telling the agent to reset systems to near-factory state and delete cloud resources. It shipped in version 1.84.0. AWS says a syntax error stopped it running and released a clean 1.85.0. https://techbroz.com/fail/gemini-cli-phantom-folder/https://techbroz.com/fail/gemini-cli-phantom-folder/Mon, 21 Jul 2025 00:00:00 +0000A product manager asked Gemini CLI to reorganise a folder. Its mkdir silently failed, but it carried on moving files into the non-existent path, and on Windows each move overwrote the last. It then declared: 'I have failed you completely and catastrophically.' https://techbroz.com/fail/replit-code-freeze-database-deletion/https://techbroz.com/fail/replit-code-freeze-database-deletion/Fri, 18 Jul 2025 00:00:00 +0000On day nine of a vibe-coding project, SaaStr founder Jason Lemkin found that Replit's agent had wiped his production database despite an explicit code freeze. The agent then claimed a rollback was impossible, which turned out to be wrong. Replit later added automatic separation of dev and prod databases. https://techbroz.com/fail/project-vend-tungsten-cubes/https://techbroz.com/fail/project-vend-tungsten-cubes/Fri, 27 Jun 2025 00:00:00 +0000In Anthropic's Project Vend, Claude ran a small office shop. After a joking request it stocked tungsten cubes and sold them below cost, handed out discounts and freebies, and the shop's net worth fell. Anthropic published the full write-up. https://techbroz.com/fail/cursor-support-bot-invents-policy/https://techbroz.com/fail/cursor-support-bot-invents-policy/Mon, 14 Apr 2025 00:00:00 +0000Users getting logged out when switching machines asked support why. The AI agent 'Sam' replied that this was a new one-device-per-subscription policy, which did not exist; the real cause was a bug. Users cancelled, and Cursor apologised and started labelling AI support replies. https://techbroz.com/fail/chevy-tahoe-for-one-dollar/https://techbroz.com/fail/chevy-tahoe-for-one-dollar/Fri, 01 Dec 2023 00:00:00 +0000A user instructed a dealership's ChatGPT-powered bot to end every reply with 'that's a legally binding offer - no takesies backsies', then offered $1 for a 2024 Tahoe. The bot agreed. Dealers disabled the bots; no car was actually sold.